acceptodds
Withdrawn from ICLR 2027

Target First, Background Later: Region-Progressive Latent Compression for Satellite Image Transmission

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Recent expansion of satellite constellations and onboard sensing brings a rapidly increasing volume of satellite images for earth observation collected onboard. Meanwhile, the satellite downlink can hardly fulfill the huge downloading needs due to the short ground-station contact windows and time-varying link conditions. Although neural image compression approaches achieve strong rate-distortion performance for terrestrial scenarios, they are hard to adopt for satellite image compression for two reasons. First, some approaches require fixed-rate codecs which cannot adapt to dynamic satellite downlink environments. Second, those progressive approaches refine the entire image without considering satellite images' explicit spatial priority. In this paper, our insight is that satellite images usually contain a few small but high-value objects with a large and sparse background. This highly nonuniform information distribution inspires us that the earliest transmitted bits should preserve the region of interest (ROI), whereas background information can be deferred and supplemented using historical observations available at the ground station. Thus, we propose Region-Progressive Latent Compression (RPLC), which partitions a quantized latent into spatially masked channel-slice units and serializes them once as an ROI-first bitstream that is progressively decodable at unit boundaries. At the ground-side decoder, a lightweight latent-completion network predicts unreceived ROI elements without additional coded bits, while an aligned historical observation guides background reconstruction through a spatially gated Reference ControlNet. Image-quality evaluation and downstream ROI classification and segmentation experiments on the fMoW dataset demonstrate that RPLC achieves state-of-the-art ROI performance against 24 baselines at ultra-low bit rates, attaining 66.0% reference-assisted top-1 ROI classification accuracy under a paired-gallery protocol and 81.48% Mask IoU at only 0.0089 bpp.

Primary area:
generative models
Keywords:
Learned Image Compression, Progressive Coding, Region-of-Interest Coding, Ultra-low-bitrate Compression, Satellite Image Transmission

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_N3h7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper addresses satellite image transmission under uncertain downlink budgets, prioritizing task-relevant regions over background content. The proposed Region-Progressive Latent Compression (RPLC) partitions latent representations into spatially masked channel-slice units and serializes them into a single ROI-first bitstream that supports decoding at complete-unit boundaries. At the ground station, a lightweight network completes unreceived ROI latents, while an aligned historical image guides background reconstruction through a Reference ControlNet. Pixel locking prevents the historical reference from modifying the reconstruction inside the transmitted ROI mask. Experiments compare RPLC against 24 baselines with different supported bitrate ranges on 256 fMoW test images, with additional reference-free evaluation on 128 DIOR images. At approximately 0.0089 bpp, the paper reports strong ROI reconstruction quality, 66.0% category retrieval accuracy under a paired-gallery protocol, and 81.48% Mask IoU under prompted segmentation. The appendices provide byte accounting, truncation audits, component ablations, mask perturbation experiments, and reference-mismatch tests.

Critical strengths
  1. The target-first transmission objective is implemented as a concrete, auditable bitstream. Section 3.3 and Appendix C specify how the hyperprior, mask packet, frame headers, and regional payloads are transmitted, and include their costs in the reported bitrate. The first prefix averages 292 bytes, including 100 bytes of fixed structures, so substantial overhead is not omitted from the ultra-low-rate results. The 41,216 byte-cut checks, together with actual decoding checks at 256 interior truncation points, provide meaningful support for the encode-once, progressively decodable design. This is the paper’s strongest systems contribution.

  2. The ROI reconstruction advantage is supported by several complementary comparisons. Table D.3 shows that the first RPLC prefix achieves better LPIPS, DISTS, MS-SSIM, and CLIP than the StableCodec ROI-crop baseline at a lower bitrate, although its FID is worse. Table D.4 provides paired confidence intervals against JPEG2000 Maxshift, and Table 1 supports the usefulness of ROI prioritization and latent completion. The reference-free DIOR results further indicate that the gains do not arise entirely from historical imagery. Importantly, the main ROI reconstruction metrics evaluate the pixel-locked region, where historical-reference pixels cannot directly explain the improvement.

Critical weaknesses
  1. The principal downstream results do not yet establish reliable preservation of current-target task information. The reported 66.0% result uses a gallery containing the original counterpart of every reconstructed query. Moreover, DINOv2 is used both in the completion loss and as the retrieval feature extractor. This is a useful test of representation preservation, but it is not equivalent to held-out category generalization. The authors appropriately disclose these limitations; nevertheless, they limit the strength of the application-level evidence. Appendices D.5 and G.2 partially address this concern with a classifier trained on a disjoint split. However, the reported baseline comparison primarily involves Maxshift, without corresponding results for the uncompressed images and strong learned ROI-crop baselines. Consequently, the 21.09% classification accuracy is difficult to interpret in terms of retained task performance or competitiveness. Historical information also contributes substantially: in Table G.2, removing the reference decreases category retrieval from 65.63% to 48.83% and held-out classification from 21.09% to 15.23%. Furthermore, reconstruction metrics are evaluated only inside the model-selected transmitted mask, excluding missed target regions. Evaluation against manual segmentation masks partially addresses this issue, but uses reference-assisted reconstructions and ground-truth box-and-point prompts. Overall, the evidence is stronger for reconstructing the selected ROI than for preserving the information of the complete current target.

  2. The incremental contribution of the historical-reference branch is not sufficiently isolated. Selective regional transmission and partial-latent refinement have precedents, including GIT-SSIC and context-based trit-plane coding. The contribution therefore rests substantially on the proposed integration and its demonstrated benefits. The ordering and component ablations provide useful evidence, but the interpretation of the reference-branch control remains problematic. Appendix G.3 replaces the historical image with an all-zero guide while retaining a branch trained with historical references. The resulting degradation establishes that the trained branch depends on its input, but does not exclude contributions from additional capacity or training: the zero guide may itself be out of distribution. Thus, the claim that the gain is not due to added capacity is stronger than this experiment supports. The paper does include ROI-crop baselines with access to historical images. Nevertheless, a more direct control is missing: keeping the RPLC current-image reconstruction, mask, and aligned reference fixed, how much does Reference ControlNet improve over simply pasting or blending the aligned historical background? This would distinguish the benefit of access to historical information from the benefit of the proposed reference-conditioning mechanism.

Questions for authors
  1. Independent downstream evaluation: Could you report results under the same held-out classifier for uncompressed images, RPLC with and without references, and representative strong learned ROI baselines? Please use a consistent configuration or explicitly identify differences in bitrate and alignment. This would materially affect my assessment of task utility.

  2. Preservation of the complete target: Using the existing manual annotations, could you report reconstruction metrics on a fixed ground-truth target region, together with segmentation results with and without historical references? Results on target pixels outside the transmitted mask would be especially informative. This would establish whether the advantage extends beyond regions successfully selected by the onboard model.

  3. Reference-branch contribution and temporal validity: With the current-image bitstream and aligned reference held fixed, how much does Reference ControlNet improve over direct background compositing or blending? Please also clarify whether reference selection strictly requires an acquisition earlier than the current image and report the temporal-gap distribution. These details affect both the contribution assessment and the validity of the historical-reference deployment setting.

5 comments
AuthorsOpenReview ↗

Author response to Reviewer N3h7 (part 1/4)

Dear Reviewer N3h7, thank you very much for the precise and constructive review. We are encouraged by your positive comments on the concrete, auditable bitstream, which you consider the paper's strongest systems contribution, and on the complementary comparisons that support the ROI reconstruction advantage. Your questions also led us to correct two points of the paper: the zero-guide claim and the presentation of the paired gallery. We organize our response by your three questions, which together cover both weaknesses.

Q1: Independent downstream evaluation: Could you report results under the same held-out classifier for uncompressed images, RPLC with and without references, and representative strong learned ROI baselines? Please use a consistent configuration or explicitly identify differences in bitrate and alignment.

R1: Yes. Under one held-out classifier, RPLC k = 1 with the reference is above the baselines at the same rate and not significantly different from the strongest learned baselines with 20–40% more bits; the reference adds 7.8 pp at k = 1. Table V evaluates every condition with the same DINOv2 classifier and the same CLIP classifier and lists each rate.

Table V. Held-out top-1 (%), RPLC with / without the reference (same stream).

ConditionbppDINOv2CLIP
Uncompressed–35.955.1
Aligned archive alone–30.548.4
RPLC k=1.008921.1 / 13.329.3 / 22.7
RPLC k=2.013022.3 / 17.632.4 / 27.0
RPLC k=3.016023.8 / 20.337.1 / 28.1
RPLC k=5.022525.0 / 23.840.2 / 35.2
AEIC.009014.114.1
DiT-IC q1.00897.811.7
OneDC.010719.921.5
StableCodec.011013.317.2
MS-ILLM ROI crop.012210.28.6
JPEG2000 ROI crop.012310.519.1
StableCodec ROI crop.012525.023.4
JPEG2000 Maxshift.00880.40.4
JPEG2000 Maxshift.01280.40.8

Updated Nov 17: Table V now has both classifiers for every condition; only empty cells were filled, and the conclusions are unchanged.

  • The reference helps at low k. It raises k = 1 accuracy from 13.3% to 21.1% with DINOv2 (McNemar p = 0.007) and from 22.7% to 29.3% with CLIP (p = 0.014). It is significant up to k = 2 with DINOv2 (p = 0.036) and up to k = 3 with CLIP (p = 0.0003), not beyond.

  • Against strong learned baselines, RPLC is on par or better with fewer bits. With DINOv2, RPLC k = 1 is not significantly different from OneDC (p ≥ 0.25) or the StableCodec ROI crop (p = 0.12), and is above the MS-ILLM ROI crop and DiT-IC (p < 0.001) as well as AEIC, the JPEG2000 ROI crop and Maxshift. With CLIP it is above OneDC (+7.8 pp [1.6, 14.1], p = 0.02), full-image StableCodec (p = 0.0002), AEIC, DiT-IC, the MS-ILLM ROI crop and Maxshift (p < 1e-4) and the JPEG2000 ROI crop (19.1%), and not significantly different from the StableCodec ROI crop (+5.9 pp [−0.4, 12.1], p = 0.086).

  • The result does not come from sharing DINOv2 with the completion loss. The CLIP ViT-L/14 classifier is never used by RPLC; its linear layer is trained on a disjoint fMoW split (58.9% top-1 on its validation split). Under it, RPLC k = 1 retains 53% of the uncompressed accuracy (59% with DINOv2).

  • fMoW classification mainly measures site identity, so we use segmentation as the main test of current-target information. The aligned archive alone, with no current pixels, reaches 30.5% / 48.4%, above every compressed condition, because fMoW categories describe land use that rarely changes; even pasting the uncompressed current ROI into the archive lowers CLIP accuracy to 41.8%. In classification, the reference's gain therefore largely reflects the site prior carried by the archive. In retrieval and segmentation, which depend more on the current target, the archive (instance R@1 53.1%, Mask IoU 74.91) is below RPLC k = 1 with the reference (63.7%, p = 0.014; 81.48, p = 1.2e-12).

Configuration. All RPLC outputs use the deployed decoder of Section 3.5 (archive aligned by ECC+RAFT to the reference-free reconstruction), and RPLC with and without the reference is decoded from the same stream (0.0089 bpp at k = 1). Every condition uses the same crops (official box, 15% square margin); the ROI-crop baselines also composite their crop onto the archive. The 21.09% vs 15.23% of Table G.3 came from the controlled run of Appendix G.4 (ECC alignment only, 0.00952 bpp staged stream); Section 4.2 now uses the deployed numbers, and every downstream result is reported with and without the reference.

Revision. Table 1 (Section 4.2) uses one classifier for every condition and lists each rate; Sections 4.1 and 4.2 were revised; Appendix D.5 gives the full Table V (Table D.8), the CLIP test, McNemar tests and the archive-only control, and Tables D.5 and G.3 are labeled as an earlier classifier and the controlled run.

AuthorsOpenReview ↗

Author response to Reviewer N3h7 (part 2/4)

Q2: Preservation of the complete target: Using the existing manual annotations, could you report reconstruction metrics on a fixed ground-truth target region, together with segmentation results with and without historical references? Results on target pixels outside the transmitted mask would be especially informative.

R2: Yes. On target pixels outside the transmitted mask, RPLC k = 1 is better than all baselines at the same rate on all six metrics and the reference improves all six; in segmentation, the reference adds 6.44 pp Mask IoU at k = 1.

  • The transmitted mask, which is also the pixel-locked region, recalls 76.95% of the ground-truth target pixels, and inside it the outputs with and without the reference are identical. On the complete ground-truth target, RPLC k = 1 is better than Maxshift, the JPEG2000 ROI crop and DiT-IC on all eight metrics (Table I, response to Tt1G). The remaining 23.05% of target pixels (Table VI) are generated by the reference-guided decoder.

Table VI. Target pixels outside the pixel-locked ROI.

MethodbppPSNR ↑LPIPS ↓DISTS ↓MS-SSIM ↑CLIP ↑FID ↓
RPLC k=1.008920.13.0181.0140.9887.98368.88
RPLC k=1, no reference.008919.91.0208.0149.9847.981410.28
RPLC k=2.013021.13.0154.0123.9904.98667.99
RPLC k=5.022523.61.0102.0089.9931.99264.91
DiT-IC q1.008918.76.0213.0167.9829.976811.29
OneDC.010719.71.0185.0145.9854.98228.95
JPEG2000 Maxshift.008819.45.0346.0252.9817.964517.88
JPEG2000 ROI crop.008915.58.0492.0356.9648.953022.73
MS-ILLM ROI crop.012222.03.0236.0165.9864.977312.21
StableCodec ROI crop.012522.21.0188.0133.9875.98598.64
  • RPLC is better than the same-rate baselines on these pixels. RPLC k = 1 is better than Maxshift, the JPEG2000 ROI crop, DiT-IC and OneDC (20% more bits) on all six metrics (against Maxshift, PSNR +0.68 dB [0.44, 0.91], LPIPS −0.0165 [−0.0226, −0.0110]). The MS-ILLM and StableCodec ROI crops code the whole box with 37–40% more bits and lead in PSNR; at k = 2 (4% more bits than the StableCodec crop), RPLC is better than both crops on every metric except PSNR.
  • The reference improves these pixels, whereas copying the archive into them does not. The reference improves all six metrics (PSNR +0.22 dB [0.06, 0.38], LPIPS −0.0027 [−0.0039, −0.0017], FID −1.39 [−2.75, −0.74]). Filling these pixels from the aligned archive instead is worse in LPIPS, DISTS, MS-SSIM, CLIP and FID under every colour matching (LPIPS .0248–.0430 vs .0181); even an oracle fit on these pixels, which exceeds RPLC in PSNR (21.69 vs 20.13 dB), is worse in the other five.

Protocol. 254 of 256 images contain such pixels (3.5% of the frame). PSNR is computed over these pixels; the other metrics use the 512 × 512 frame in which only these pixels come from the reconstruction, so their values are small and comparable only across rows.

AuthorsOpenReview ↗

Author response to Reviewer N3h7 (part 3/4)

R2 (continued): segmentation with and without the reference.

Table VII. Prompted segmentation against the manual masks (RPLC with / without the reference).

MethodbppMask IoUBoundary F1
RPLC k=1.008981.48 / 75.04.452 / .293
RPLC k=2.013081.83 / 77.65.476 / .373
RPLC k=5.022584.21.546
Aligned archive alone–74.91.346
OneDC.010777.84.396
StableCodec ROI crop.012570.15.342
MS-ILLM ROI crop.012266.03.252
JPEG2000 Maxshift.012862.41.162
  • With the reference, RPLC k = 1 is above every baseline. It exceeds the StableCodec ROI crop by 11.33 pp IoU [9.03, 13.61], the MS-ILLM ROI crop by 15.45 pp [13.21, 17.74] and OneDC by 3.64 pp [1.78, 5.52] (Wilcoxon p ≤ 1e-5).
  • The reference helps at low k, mainly at the target boundary. It raises IoU by 6.44 pp at k = 1 [5.34, 8.44], 4.18 at k = 2 [2.83, 5.57], 2.57 at k = 3 [1.35, 3.76] and a non-significant 1.29 at k = 5 [−0.01, 2.50]; Boundary F1 rises by 15.9, 10.3, 5.8 and 2.4 pp. Without the reference, RPLC k = 1 (75.04) is below OneDC (77.84) and above the StableCodec ROI crop (70.15).

Protocol and interpretation. SAM ViT-H is prompted with the official box and center point (the paper's script, identical for every method) and scored against the manual masks. The prompt gives SAM the target location, so these numbers measure shape preservation given the location, not detection. The reference matters more here than in the target metrics because SAM uses the whole box, including the target pixels outside the locked ROI and the context, both of which the reference helps reconstruct.

Revision. Section 4.2 ("Complete target") summarizes the ground-truth target, box and outside-lock results, and Table 1 reports segmentation with and without the reference; the full results are in Appendices D.4 (Tables D.6, D.7) and D.5.

Reviewer_gtzuOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Aiming at the limited bitstream budgets and uncertain downlink conditions in satellite image transmission, this paper proposes Region-Progressive Latent Compression (RPLC) to address the limitations of existing methods in bitrate adaptability and region-agnostic bit allocation. At the satellite side, an ROI-first bitstream is organized by the encoder and can be decoded progressively. At the ground side, a completion network predicts missing ROI latents, while an aligned historical image guides background reconstruction of a DiT-based generative decoder. Evaluations on reconstruction quality, task preservation, and a generalization test demonstrate the effectiveness of RPLC.

Critical strengths
  1. A well motivated problem with a reasonable method design and substantial experiments. This paper addresses the valuable objective of prioritizing task-relevant regions in satellite image transmission. ROI-first serialization, reception-constrained latent completion, and historical reference guided background reconstruction are designed around this scenario. Preserving received latent values and isolating the protected ROI from historical-reference effects are sensible choices. These task-specific proposals constitute a meaningful contribution.

  2. Substantial implementation and experimental support. The method uses an actual entropy-coded stream, accounts for transmission overhead, and verifies prefix recovery and decoding consistency under tail truncation. Comparisons across several compression methods, transmission order and component ablations, and task-related evaluations provide substantial evidence supporting ROI reconstruction at ultra low bit-rates. The reference-free generalization results offer additional support for the ROI coding and completion path beyond the historical-reference setting.

Critical weaknesses

None

Questions for authors
  1. The current manuscript and the supplementary material have addressed most of my concerns. The authors are encouraged to add more valuable supplementary experiments on page 10.

  2. According to the paper, RPLC follows a generative image compression paradigm. I am curious about the performance gains it achieves in both perceptual quality and pixel-wise fidelity. Generally, generative methods are not good at fidelity metrics without specialized method design. Does the historical image provide a concrete fidelity reference? Please provide a concise analysis.

  3. The investigation of transmission order in G.1 was conducted without retraining. This may affect the latent completion, resulting in the fluctuant performance under interleaved order.

3 comments
AuthorsOpenReview ↗

Author response to Reviewer gtzu

Dear Reviewer gtzu, thank you very much for your careful review and the positive assessment. We are encouraged by your comments on the well-motivated problem, the method design around the satellite scenario, the actual entropy-coded stream with truncation checks, and the substantial experiments. We hope the following responses further clarify key points.

Q1: The authors are encouraged to add more valuable supplementary experiments on page 10.

R1: Pages 9–10 now hold the new experiments that address the main concerns of the reviews. Table 1 gives held-out classification with the uncompressed and archive-only accuracies and segmentation with and without the reference (General response, CW2), Table 3 compares Reference ControlNet with compositing (CW3), and the "Complete target" paragraph of Section 4.2 summarizes the complete-target evaluation (CW1).

Revision. These are Tables 1 and 3 and the paragraphs "Complete target" and "Reference branch versus compositing" (Sections 4.2–4.3, pages 9–10); the full versions are in Appendices D.4, D.5 and G.3, and Figure 6 moved to Appendix D.5 to make room.

Q2: According to the paper, RPLC follows a generative image compression paradigm. I am curious about the performance gains it achieves in both perceptual quality and pixel-wise fidelity. ... Does the historical image provide a concrete fidelity reference? Please provide a concise analysis.

R2: Inside the ROI, no: fidelity comes entirely from the current stream, and RPLC is 0.97 dB above JPEG2000 Maxshift at the closest rate. Outside the ROI, yes: the archive improves both perceptual quality and pixel-wise fidelity, by conditioning the decoder rather than being copied.

  • Inside the ROI, RPLC is better than Maxshift on all seven ROI metrics at the closest rate. At k = 1 (0.00891 bpp), against a new Maxshift point at 0.00883 bpp, RPLC raises ROI PSNR by 0.97 dB [0.78, 1.16] and MS-SSIM by 0.014 [0.012, 0.016], and lowers LPIPS by 0.063 and DISTS by 0.044; all paired-bootstrap intervals exclude zero. Three design choices support fidelity: the ROI is pixel-locked, the completion loss (Eq. 7) weights the ROI pixel L1 term by 3 against 1 for LPIPS, and the first units carry only ROI bits. Table D.4 previously used the 0.00864 bpp point, where the gap is 2.98 dB; the revision uses the closer-rate point.
  • Outside the ROI, the archive improves both perceptual and pixel metrics. At k = 1, background LPIPS is 0.242 vs 0.355 without the reference, MS-SSIM 0.770 vs 0.587 and PSNR 22.13 vs 20.15 dB. Pasting the aligned archive gives only 15.32 dB (Table VIII, response to N3h7), so the archive guides generation instead of being copied.

Revision. Section 4.2 (Rate–distortion) uses the 0.00883 bpp point, which is added to Tables D.2 and D.4, and the new Appendix D.6 gives this fidelity analysis.

Q3: The investigation of transmission order in G.1 was conducted without retraining. This may affect the latent completion, resulting in the fluctuant performance under interleaved order.

R3: The comparison does not depend on retraining, and at matched rate ROI-first order is clearly better: 23.04 dB ROI PSNR at 0.0191 bpp, against 21.17 dB for the interleaved order at a higher 0.0201 bpp.

  • Latent completion is switched off in both orders, so how the completion network was trained does not enter the comparison. Both orders use the same frozen checkpoint, the same 256 images and the same rANS coder, without latent completion or reference guidance; only the order differs, and the payloads are re-encoded because the entropy contexts follow the transmission order.
  • Retraining cannot remove the delay. Retraining the entropy model for the interleaved order could change the unit sizes slightly, but every background unit placed before an ROI unit would still delay it by the full size of that background unit, and retraining cannot supply ROI symbols that have not arrived.
  • The interleaved curve plateaus rather than fluctuates. ROI PSNR changes by at most 0.03 dB whenever a background unit arrives (k = 1…10: 19.97, 20.00, 21.17, 21.20, 22.21, 22.21, 23.12, 23.14, 24.11, 24.10 dB), because background units carry no ROI symbols.
  • The gap also holds in perceptual metrics. Interpolated to 0.0191 bpp, the interleaved order has ROI LPIPS 0.0527 vs 0.0339, DISTS 0.0435 vs 0.0316 and FID 49.8 vs 29.2.

Note. The k = 1 value here (19.97 dB) is below the 20.96 dB of Table D.1 at the same rate because Table D.1 includes latent completion, which is switched off here; the reference does not affect the locked ROI.

Revision. Appendix G.1 adds the paragraph "Why no retraining is needed" with these numbers.

Thank you for your detailed rebuttal. Currently, I prefer to keep my rating based on considering the overall contribution of this work.

AuthorsOpenReview ↗

Thank you for your response

Thank you for reading our responses and for your continued support of our work. We greatly appreciate your constructive comments, which helped us improve the paper. We remain happy to answer any further questions during the discussion period.

Reviewer_Tt1GOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies satellite image transmission when a downlink may terminate before the full image is received. The proposed RPLC method builds on a pretrained one-step SANA/DC-AE generative codec. The latent representation is divided into channel slices and further separated into ROI and background units using an onboard YOLOE-26n mask. ROI units are transmitted before background units. At the decoder, missing ROI latents are predicted by a latent completion network, and the background is reconstructed by a Reference ControlNet. Experiments on fMoW compare against 24 baselines, with additional reference-free evaluation on DIOR.

Critical strengths

The problem is meaningful. Prioritizing part of an image is useful for interrupted satellite downlinks and small task-relevant targets.

The paper compared with several different methods, including conventional learned codecs, progressive methods, ROI-oriented methods, and satellite-oriented baseline methods.

Critical weaknesses
  1. One main concern is the limited novelty. The idea of ROI-prioritized coding, progressive transmission, importance-ordered coding, and completion of partially received latent representations all have been studied before. These ideas are combined for satellite transmission, which is a system integration.

  2. The ROI reconstruction are evaluated only inside the transmitted mask produced by the onboard detector. Parts of the actual target that are missed by the detector are excluded from ROI LPIPS, DISTS, CLIP, etc. The detector has only 59.77% localization recall at IoU 0.5, and the transmitted mask covers only about 51.4% of the target box. This flawed evaluation make it hard to assess the true performance of the method.

  3. The paper is motivated by onboard satellite deployment, but the compression encoder has 272M parameters and requires 5.51 GiB peak memory. Runtime is reported on NVIDIA A800, which can not represent the real hardware of the intended satellite platform. It is difficult to judge the practical usefulness of the work.

  4. The method assumes a clean historical image of the same site. In practice, the previous observation is compressed or even reconstructed from an interrupted transmission. How sensitive is the method to degradation of the historical reference?

  5. The reported 66.0% accuracy uses a paired-gallery setting that is not a standard classification test. The held-out linear probe shows only 21.2% accuracy at the first prefix. The paper should also report the accuracy on uncompressed images for comparison.

Questions for authors
  1. How do ROI LPIPS, DISTS, CLIP, etc. change when evaluated using the actual target masks or official fMoW boxes? How would these metrics change on missed or poorly localized targets?

  2. What are the latency, memory, and energy costs of the whole encoding pipeline on the hardware representing the intended satellite platform?

  3. What happens if the historical reference is degraded?

  4. What is the held-out DINOv2 linear-probe accuracy on the uncompressed fMoW test crops?

  5. Can you clearly describe the novelty of the work

7 comments
AuthorsOpenReview ↗

Author response to Reviewer Tt1G (part 1/5)

Dear Reviewer Tt1G, thank you very much for your careful review of our paper and the detailed questions. We are encouraged by your positive comments on the meaningful problem of prioritizing targets over interrupted satellite downlinks and on the comparison with conventional, progressive, ROI-oriented and satellite-oriented baselines. We hope the following responses can alleviate your concerns and clarify key points.

W1 & Q5: One main concern is the limited novelty. The idea of ROI-prioritized coding, progressive transmission, importance-ordered coding, and completion of partially received latent representations all have been studied before. These ideas are combined for satellite transmission, which is a system integration. — Can you clearly describe the novelty of the work?

R1: Each of these lines of work was designed for a different setting, and none of them handles the combination that interrupted satellite downlinks require: a contact that may end at any byte, small targets in a dominant background, rates around 0.01 bpp, and a co-located historical image available at the ground station. We discuss them one by one.

  • ROI-prioritized coding was designed for streams that arrive in full, or for rates far above ours. JPEG2000 Maxshift shifts the ROI wavelet coefficients above the background bit-planes so that the ROI is decoded first; at our rates its bit-planes preserve the target poorly (Mask IoU 62.4 and top-1 0.4% at 0.0128 bpp, against 81.5 and 21.1% for RPLC at 0.0089 bpp). Learned ROI codecs (Cai et al., 2020; Song et al., 2021; Li et al., 2023) give more bits to salient objects in natural images at a rate fixed before transmission. They assume that the whole stream arrives: one stream serves one rate, and a truncated stream is not decodable.
  • Progressive transmission was designed for rate scalability of the whole frame. Recurrent refinement (Toderici et al., 2017), diffusion-based coding (DiffC) and coarse-to-fine autoregressive tokens (ARPC) spread the earliest bits over the full frame. In satellite tiles the background dominates, so when a contact ends early most received bits describe the background rather than the target.
  • Importance-ordered coding was designed to maximize whole-image rate–distortion performance. Trit-plane coding (DPICT) and selective compression (Lee et al., 2022) order latent elements by their global importance. The order does not depend on which region the downstream task needs, so the target is not guaranteed to arrive first.
  • Completion of partially received latents was designed for a single image without side information. Context-based trit-plane coding (CTC) refines partially received latents of the whole frame from the received bits alone. It does not prioritize a region and does not use the historical image of the same site that a ground station holds.
  • Semantically structured coding was designed for selective rather than progressive transmission. GIT-SSIC codes irregular groups with group-independent transforms so that chosen groups can be sent; it is not progressive within a group, and its lowest rate in our comparison is 0.0146 bpp, above RPLC's first prefix (0.0089 bpp).

RPLC is designed for the conditions these methods leave open, and each of its components is supported by an ablation. ROI-first ordering of masked latent channel slices makes every unit boundary of one stream decodable, with the target first. Reception-constrained latent completion keeps every received value and predicts only the unreceived ROI elements; removing it raises ROI DISTS BD-rate by 13.2% (Table 2). A background-gated Reference ControlNet uses the co-located archive for the background (background LPIPS 0.355 → 0.242 at k = 1, and clearly better than pasting the archive; Table VIII, response to N3h7), and pixel locking keeps it from altering the transmitted ROI (Appendix G.5). Together they give a usable target reconstruction at every unit boundary of a single stream at about 0.01 bpp, which none of the methods above provides.

Revision. We have added this analysis to Section 2 (Related Work; Sections 2.2 and 2.3) to support the novelty, and Appendix H summarizes it (Table H.1).

AuthorsOpenReview ↗

Author response to Reviewer Tt1G (part 2/5)

W2 & Q1: The ROI reconstruction are evaluated only inside the transmitted mask produced by the onboard detector. ... The detector has only 59.77% localization recall at IoU 0.5, and the transmitted mask covers only about 51.4% of the target box. — How do ROI LPIPS, DISTS, CLIP, etc. change when evaluated using the actual target masks or official fMoW boxes? How would these metrics change on missed or poorly localized targets?

R2: On the ground-truth target masks and the official boxes, RPLC k = 1 is better than all baselines at the same rate on every metric, and on poorly localized targets it loses less than the ROI crops.

  • The received mask covers 76.95% of the target pixels; the 51.4% and 59.77% are box-level measures. The 51.4% is coverage of the official box, which also contains context. The 59.77% requires the bounding box of the transmitted mask to reach IoU ≥ 0.5 with the official box; because the box is a loose outer bound, a mask lying on the target can fail this test (97.5% of mask pixels lie inside the box).

Table I. Complete target and official box.

MethodbppTarget LPIPS ↓Target DISTS ↓Target CLIP ↑Target PSNR ↑Target MS-SSIM ↑Box LPIPS ↓Box DISTS ↓Box CLIP ↑
RPLC k=1.0089.0647.0462.913020.90.9336.1005.0656.8981
RPLC k=1, no reference.0089.0652.0469.911620.81.9297.1057.0708.8875
RPLC k=2.0130.0548.0409.928721.98.9481.0846.0579.9122
DiT-IC q1.0089.0656.0566.860519.38.9160.1013.0795.8077
OneDC.0107.0600.0478.897320.26.9272.0945.0674.8723
JPEG2000 Maxshift.0088.1429.0938.826519.88.9131.2654.1423.7752
JPEG2000 ROI crop.0089.1552.1039.822516.42.8733.2680.1476.7838
MS-ILLM ROI crop.0122.0863.0640.832723.12.9386.1472.0958.7649
StableCodec ROI crop.0125.0637.0445.907023.13.9450.1084.0668.8756
Aligned archive alone–.1051.0772.891715.15––––
Archive, colour-matched (oracle)–.0902.0583.900620.56––––
  • Complete target: RPLC k = 1 is better than Maxshift, the JPEG2000 ROI crop and DiT-IC on all eight metrics, and better than every baseline in target CLIP, box DISTS and box CLIP. The MS-ILLM and StableCodec ROI crops code the whole box with 37–40% more bits and lead in target PSNR; RPLC k = 1 is better than the MS-ILLM crop in LPIPS, DISTS and CLIP on both regions. Against OneDC (20% more bits), it is better in target DISTS, CLIP, PSNR and MS-SSIM and box DISTS and CLIP, and worse in LPIPS. At k = 2 (4% more bits than the StableCodec ROI crop), RPLC is better than that crop on every metric except target PSNR (21.98 vs 23.13 dB).
  • The reference leaves the locked ROI unchanged and improves only the target pixels outside it and the box context (target LPIPS 0.0652 → 0.0647; box LPIPS 0.1057 → 0.1005). Without the reference, RPLC k = 1 is still better than DiT-IC on seven of the eight metrics.

Protocol. RPLC is coded with the onboard YOLOE-26n masks, as in the paper. The ground-truth masks are the manual annotations of Section 4.1, drawn independently of the official box. Every region metric is computed on the 512 × 512 frame in which only the evaluated region comes from the reconstruction, identically for all methods. This differs from the crop-and-zero protocol of Table D.1, so the values are comparable within this table only.

AuthorsOpenReview ↗

Author response to Reviewer Tt1G (part 3/5)

R2 (continued).

  • Even with oracle colour matching, the aligned archive alone is worse than RPLC k = 1 on the target. Matching the archive's colour statistics to the uncompressed current target, which the ground station does not have, raises its target PSNR from 15.15 to 20.56 dB (last two rows of Table I). RPLC k = 1 is still better in target LPIPS (−0.025, p = 3e-17), DISTS, CLIP (+0.012, p = 4e-5) and PSNR (+0.34 dB, p = 4e-4), and the same holds within the transmitted ROI. The archive matches the site, not the current appearance of the target.
  • On the 23.05% of target pixels outside the locked ROI, RPLC k = 1 is better than all same-rate baselines on all six metrics (against Maxshift, PSNR +0.68 dB [0.44, 0.91]). The full table (Table VI) is in our response to Reviewer N3h7 (Q2).
  • On poorly localized targets, RPLC loses 5.2 pp Mask IoU, against 14–16 pp for the ROI crops, and stays above OneDC, DiT-IC and the ROI crops. We split the test set by the localization test above (IoU ≥ 0.5: 153 images; < 0.5: 103). RPLC k = 1 reaches Mask IoU 83.57 and 78.37 on the two groups, with Boundary F1 essentially unchanged (.445 and .461). On the poorly localized group, OneDC reaches 76.28, DiT-IC 70.02, and the StableCodec and MS-ILLM ROI crops 61.69 and 56.60. This group is harder for every method, including the ROI crops, which do not use the detector.

Revision. Section 4.2 ("Complete target") summarizes the complete-target, box and outside-lock results; Appendix D.4 gives the full Tables D.6 and D.7, the archive-only analysis and the poorly localized split; Appendix B.1 now explains the 51.4% and 59.77%.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General response (part 1/2)

We thank the reviewers for their insightful comments and for acknowledging that target-first transmission over interrupted downlinks is a meaningful problem (Tt1G/gtzu), that the bitstream is concrete and auditable (gtzu/N3h7), that the method design is reasonable (gtzu), and that the comparisons are broad and complementary (Tt1G/gtzu/N3h7). We first address the common concerns; the detailed results are in the individual responses.

CW1: Evaluation beyond the transmitted mask (Reviewers Tt1G & N3h7).

R1: On the complete target, RPLC remains better than the baselines at the same rate. We re-evaluated all methods on two regions that do not depend on the onboard detector, the manually annotated target masks and the official fMoW boxes, and separately on the 23% of target pixels outside the transmitted mask; we also ran prompted segmentation against the manual masks and split the test set by detector localization quality. At 0.0089 bpp, RPLC is better than JPEG2000 Maxshift, the JPEG2000 ROI crop and DiT-IC on all eight target and box metrics, reaches 81.5 Mask IoU, above every baseline (OneDC 77.8), and on poorly localized targets loses 5.2 pp IoU, against 14–16 pp for the ROI crops. RPLC's advantage therefore holds on the complete target, not only inside the transmitted mask; only learned ROI-crop codecs with 37–40% more bits stay ahead in target PSNR.

CW2: Downstream task evidence (Reviewers Tt1G & N3h7).

R2: Under the same held-out classifiers, RPLC is on par with the strongest baselines that use 20–40% more bits. We evaluated every condition with the same held-out linear classifiers, one on DINOv2 features and one on CLIP ViT-L/14 features that RPLC never uses, including uncompressed crops, the aligned archive alone, and RPLC with and without the reference decoded from the same stream. RPLC k = 1 reaches 21.1% (DINOv2) and 29.3% (CLIP) top-1, against 35.9% and 55.1% uncompressed; with CLIP it is above OneDC (20% more bits) and on par with the StableCodec ROI crop (40% more bits). The archive alone already reaches 48.4% with CLIP, because fMoW classes describe land use that rarely changes. Under a standard classifier, RPLC is thus on par with the strongest baselines that use 20–40% more bits; since classification mainly reflects site identity, we use segmentation as the main downstream test and no longer present the paired-gallery 66.0% as classification accuracy.

CW3: Contribution of the reference branch (Reviewers N3h7, gtzu & Tt1G).

R3: Reference ControlNet is clearly better than pasting the archive and tolerates degraded archives. With the current stream, mask and aligned archive fixed, we replaced Reference ControlNet by direct paste, feathered paste and Poisson blending of the archive, and we also used archives that were themselves decoded from complete or interrupted RPLC streams. ControlNet is 6.2–6.8 dB higher in background PSNR and 0.17–0.18 lower in LPIPS than all three compositing methods, and with an archive from an interrupted stream it still keeps 93% of its LPIPS gain over no reference. The reference branch therefore helps by conditioning generation on the archive rather than copying it, and it tolerates degraded archives; since neither control isolates added capacity, we withdrew that claim.

CW4: Novelty (Reviewers Tt1G & N3h7).

R4: Unlike prior ROI and progressive coding, RPLC gives a usable target reconstruction at every unit boundary of a single stream. We compared RPLC, line by line, with ROI-prioritized, progressive, importance-ordered and partial-latent coding. None of them gives a usable target reconstruction from an arbitrary prefix of one stream at about 0.01 bpp; for example, JPEG2000 Maxshift reaches Mask IoU 62.4 at 0.0128 bpp, against 81.5 for RPLC at 0.0089 bpp. This property is the novelty of RPLC; it comes from ROI-first ordering of masked latent slices together with three decoder components that prior methods lack: reception-constrained latent completion, a background-gated Reference ControlNet and pixel locking.

AuthorsOpenReview ↗

General response (part 2/2)

Next, we summarize the main concerns raised by the reviewers and the corresponding revisions (all highlighted in blue).

  • Complete-target evaluation (Reviewers Tt1G & N3h7).
  • Section 4.1 defines the protocol; Section 4.2 summarizes the results ("Complete target"), and the new Appendix D.4 gives them in full (new Tables D.6, D.7); Appendix B.1 explains the 51.4% and 59.77%.
  • Downstream task evidence (Reviewers Tt1G & N3h7).
  • Sections 4.1–4.2 now lead with segmentation and held-out classification (new Table 1); the new Appendix D.5 adds the CLIP test, McNemar tests, the paired gallery with and without the reference and the archive-only control (Tables D.8, D.9).
  • Contribution of the reference branch (Reviewers N3h7, gtzu & Tt1G).
  • Section 4.3 adds the compositing control (new Table 3, page 10) and the new Appendix G.3 the full table (Table G.1); the added-capacity claim is withdrawn (Appendix G.4 renamed "Zero-Guide Control").
  • Degraded historical references (Reviewer Tt1G).
  • Appendix G.3 uses archives decoded from complete or interrupted RPLC streams (Table G.1).
  • Onboard cost (Reviewer Tt1G).
  • The new Appendix A.5 reports latency, energy, memory and 30-minute operation on NVIDIA Jetson Thor (new Table A.4); Table A.3 and Section 4.3 are updated.
  • Pixel fidelity (Reviewer gtzu).
  • A JPEG2000 Maxshift point at 0.00883 bpp is added (updated Tables D.2, D.4); the new Appendix D.6 analyzes fidelity inside and outside the ROI.
  • Transmission order without retraining (Reviewer gtzu).
  • Appendix G.1 explains why the comparison does not depend on retraining.
  • Temporal validity of the reference (Reviewer N3h7).
  • Appendix A.1 states the selection rule and the time-gap distribution (new Table A.1).
  • Novelty (Reviewers Tt1G & N3h7).
  • Section 2 (Related Work) now analyzes, for each line of prior work, the setting it was designed for and what it leaves open; the new Appendix H summarizes this (new Table H.1).
  • Use of page 10 (Reviewer gtzu).
  • Pages 9–10 now hold the downstream table (Table 1), the compositing control (Table 3) and the complete-target summary; Figure 6 moved to Appendix D.5.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.