acceptodds
Under review as a conference paper at ICLR 2027

FastSAM2: Fully Convolutional Networks for Fast Segment Anything in Images and Videos

Abstract

Segment Anything Model 2 (SAM 2) provides a unified interface for promptable image and video segmentation, but repeated high-resolution encoding andglobal memory attention remain costly. We investigate whether a deliberatelylocal, content-conditioned memory operator can preserve quality when cross-frame correspondence is predominantly local, while trading flexible long-rangeretrieval for lower latency. We present FastSAM2, a unified fully convolutional image-video model whose Cross-frame Convolutional Aggregation (CCA) module combines temporal mixing, similarity gating, large-kernel depthwise propagation, memory-entry weighting, and gated fusion without forming a global query-key matrix. The model is trained only on public SA-1B and SA-V annotations,without pseudo-labels or distillation. Under a matched 1024 x 1024, batch-one,memory-write-inclusive protocol, FastSAM2 reaches 51.7 AP on LVIS and 132.5FPS on H20, versus 41.6 FPS for SAM 2 B+; its 3.18-3.20x speedup transfersacross three GPUs. FastSAM2 remains within 0.2-0.3 points of SAM 2 B+ onfour short/medium-duration video benchmarks, but trails SAM 2.1 B+ on SA-V(74.9 vs. 78.2) and long-term LVOS v2 (72.9 vs. 78.2). The resulting systemoffers a favorable measured efficiency-quality operating point, while its degra-dation under long absence and reappearance is consistent with the limitations oflocal spatial interaction.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Promptable Visual Segmentation, Unified Image and Video Segmentation, Fully Convolutional Networks, Cross-frame Convolutional Aggregation, Real-time Inference

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_xy4pOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies efficient promptable video object segmentation. FastSAM2 replaces global cross-frame memory attention with local convolutional aggregation and uses a shared encoder and decoder for image segmentation. Experiments cover video tracking benchmarks and image segmentation on LVIS. The method is faster with competitive results on short and medium videos, while long-term tracking remains weaker.

Critical strengths
  1. CCA is a clear technical contribution that uses local convolutions to aggregate cross-frame evidence without global memory attention.

  2. The evaluation covers image segmentation and several video benchmarks with comparisons to relevant baselines.

  3. The paper is easy to follow, and the tables and figures are informative.

Critical weaknesses
  1. CCA retrieves memory only near the object's current location, leaving long occlusions and distant reappearances without a demonstrated recovery mechanism. Speed gains on shorter sequences do not address this limitation.

  2. Recent long-video methods receive limited coverage in the experiments. SAM2Long is cited but omitted from the long-video comparison, leaving the positioning against current long-video solutions unclear.

  3. The core CCA design lacks an intuitive diagram, so its key operations are mainly conveyed through equations and text.

Questions for authors

The technical contribution is clear, but the method appears less suited to long and complex video scenarios. For the specific point, please refer to Weaknesses.

2 comments
AuthorsOpenReview ↗

Response to Reviewer xy4p

We thank the reviewer for emphasizing long-term re-identification and for suggesting a direct comparison with SAM2Long. We agree that local memory interaction is limited after long absence and distant reappearance.

1. Long-term disappearance and reappearance.(W1)

With L=3 and k=9, CCA has a 25×25 receptive field. It supports local correspondence but not global search. We re-evaluated baselines with one unified LVOS v2 val evaluator (140 videos, 239 objects). The original FastSAM2 score of 72.9 used the standard SAM 2 VOS evaluator. We have reported the shared-evaluator scores in a separate table in the paper.

MethodOriginal full-set J&FShared evaluator, full-set J&FReappearance J&F, absent ≥1 s / ≥5 sA800 frames/s (FP32)
FastSAM272.976.959.7 / 52.394.9
SAM 2 B+75.879.861.0 / 55.229.7
SAM 2.1 B+78.283.061.4 / 62.427.3
SAM2Long (SAM 2.1 B+)85.2 (repository)85.069.0 / 72.2—

The shared evaluator scores all four methods on LVOS v2 val (140 videos, 239 objects, 66,056 frames). Each object receives its ground-truth mask at first appearance. J&F is evaluated on subsequent frames except the last. An absent frame scores 1 only if the prediction is empty. The reappearance scores start at the first return after a qualifying absence (86 objects in 62 videos for ≥1 s and 27 objects in 23 videos for ≥5 s, at 6 fps). We have documented the lower reappearance scores for FastSAM2 and identified adaptive global retrieval as future work. See the main-text results (Line 371-374), the unified evaluation and subset table (Line 1298-1319), and the discussion of the reappearance gap (Line 1327-1330).

2. Comparison with SAM2Long.(W2)

We agree that citing SAM2Long without a quantitative comparison leaves the long-video positioning incomplete. We have evaluated the official SAM2Long implementation on LVOS v2 val with the official SAM 2.1 Hiera-B+ checkpoint, 1024×1024 inputs, and the first-appearance ground-truth mask as the only prompt for each object. We used three pathways, an IoU threshold of 0.1, and an uncertainty parameter of 2. We have reported 85.0 on the full set with the shared evaluator, versus 83.0 for SAM 2.1 B+ (reported 95% CI for the difference [0.0, +4.2]). The repository reports 85.2. We have given the 20.5 frames/s result separately because it comes from a bf16 end-to-end run on one 175-frame video with one object. The full-set and subset results are at Line 1309-1319, and the SAM2Long settings and timing are at Line 1320-1326.

3. CCA visualization.(W3)

We have added a simplified diagram to the main paper showing current-frame and memory inputs, temporal mixing, same-position similarity gating, local depthwise propagation, weighting across memory entries, and gated residual fusion. Section 3.3 introduces the diagram (Line 214-215), and the diagram and caption appear at Line 216-232. We have kept the detailed stage equations in the appendix (Line 760-782).

We have positioned CCA as an efficient local temporal-memory operator and identified adaptive global retrieval as a direction for future work (Line 421-425).

Thank you again for your time and effort in reviewing our submission.

We would like to kindly follow up on our rebuttal, as the rebuttal period is approaching its deadline. We have carefully addressed each of your concerns point by point in our response, and would greatly appreciate it if you could take a moment to review our rebuttal and share any further thoughts or clarifications you may have.

We would be happy to address any remaining concerns or questions.

Thank you very much for your time and consideration!

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

FastSAM2 is a unified image and video object segmentation system that replaces attention in the encoder and video memory path with convolutional components. Its Cross-frame Convolutional Aggregation (CCA) module uses local propagation, similarity gating, and weighted memory fusion. The paper reports roughly 3.2× higher processing speed than SAM2 B+ under its matched timing setup, with similar quality on four short or medium-duration benchmarks. This is a meaningful efficiency result for the combined system. However, underspecified evaluation protocols and a mismatch between the stated CCA operator and its implementation limit confidence in the broader claims.

Critical strengths
  1. The paper investigates a concrete trade-off in video object segmentation: whether a locality prior can reduce the cost of memory interaction while preserving quality in sequences dominated by nearby frame-to-frame correspondence. This is a useful question for efficient segmentation research, and the paper explicitly identifies long-range retrieval as a regime where the proposed prior may fail.
  2. The system-level result is promising. Under mentioned settings, FastSAM2 achieves a consistent speedup over SAM2 B+ while retaining competitive quality on MOSE, DAVIS 2017, YouTube-VOS 2019, and SA-V. The inclusion of memory writes in the timed region makes the video result more relevant to streaming inference. If the protocols are confirmed, this is a substantive practical contribution.
  3. The evaluation includes efficient baselines, component ablations, and a long-term benchmark rather than relying only on favorable short-video results. In particular, the weaker LVOS v2 performance provides evidence about the cost of restricting spatial interaction. The authors generally present the reported gains as properties of the combined architecture and training recipe, which is more defensible than attributing them entirely to CCA.
Critical weaknesses
  1. The LVIS comparison and image timing lack a complete task definition. Table 1 reports LVIS AP without specifying how promptable methods obtain proposals and confidence scores required by LVIS mask AP. Besides, the same table reports 41.7 image FPS for FastSAM2, whereas Table 2 reports 132.5 video FPS for that model, despite the stated image path bypassing CCA and memory. Different prompt counts or output pipelines might explain this, but the paper must disclose them.
  2. The benefit of CCA over alternative memory operators is not isolated. Table 5 removes CCA components, and Table 10 compares convolution with attention on a synthetic task. Neither replaces CCA within the same FastSAM2 system. The end-to-end results establish a useful system-level speed-quality trade-off, but do not prove that CCA itself is preferable to another local operator.
  3. Some consequential hyperparameters remain insufficiently examined. the value of the object-pointer EMA coefficient in Eq. 4 is not specified or tested, despite the discussion of long-term reappearance. The 160-dimensional memory compression is also held fixed.
  4. The depth ablation has a configuration inconsistency. Section C describes FastSAM2-lite as using L=2 with Table 2 reporting 34.5M parameters and 81.7G FLOPs. Table 9 assigns those same figures to L=3, while Section 4.3 says the ablations use the default FastSAM2 configuration, reported as 76M and 198.2G. The paper needs to identify the architecture used for each Table 9 row.
  5. Table 10’s synthetic attention comparison does not establish how CCA compares with local-window attention under the same encoder, training data, and compute budget on the paper’s main tasks.
  6. The theoretical argument has limited bearing on deployed CCA. Theorem 1 concerns an arbitrarily large convolutional family with full-domain receptive field; it provides no approximation bound for the final model. Its proof sketch also treats the image of a bounded set as compact, a repairable mathematical error. The theorem should be framed strictly as background motivation.
Questions for authors

1.The authors should specify how LVIS proposals, prompts, and confidence scores are generated, what operations and prompt counts are included in image timing, and why FastSAM2 reports 41.7 image FPS but 132.5 video FPS at the stated resolution.

2.The authors should state whether CCA and local-window attention were compared within the same FastSAM2 system on SA-V or LVOS v2. If no matched comparison exists, Table 10 should be presented as evidence about locality on a synthetic task, not CCA-specific superiority.

3.The authors should identify the backbone, CCA dimension, memory-bank size, and layer count behind Table 9’s 34.5M/81.7G row, reconcile it with the L=2 lite specification, and report the EMA coefficient and how the 160-dimensional memory compression was selected.

3 comments
AuthorsOpenReview ↗

Response to Reviewer waFd (1/2)

We thank the reviewer for highlighting the need for clearer evaluation protocols and operator-level evidence. We have revised the corresponding claims and reported the evaluation protocols alongside the results.

1. LVIS protocol.(Q1)

The released script uses no detector or automatic mask generator. It predicts one mask per ground-truth annotation, prompted by that annotation's GT box; the GT category_id is copied to the prediction. The score is the sigmoid object-score logit, and the mask is thresholded at 0.5. We use the official LVIS segmentation evaluator (300 detections/image; IoU 0.50:0.05:0.95). This is oracle box- and category-conditioned mask evaluation, not standard automatic instance segmentation. We have relabeled Table 1 as box-prompted mask AP, restricted the table to the FastSAM2 results under this documented protocol, and removed the direct comparison with detector-based AP. See the box-prompted AP table (Line 270-276) and the evaluation protocol (Line 309-314).

2. Image/video timing.(Q1)

The reported 41.7 image FPS and 132.5 video FPS are not directly comparable without specifying workload and units. Image timing must report prompts per image and whether prompts are decoded jointly; video timing is steady-state frames/s for one prompted object after memory initialization. We have reported the following timings.

ModePrompts/objectsGPU / resolutionTimed regionLatency / throughput
Image1 prompt/imageA800/1024²Full image path (encoder run once)7.4 ms / 135.1 images/s
Video1 objectA800/1024²Subsequent frames; includes memory read/write10.5 ms / 95.2 frames/s

To reconcile the A800 measurements with the H20 results in the paper: the 41.7 image FPS in the paper was measured on H20 with multiple prompts per image, whereas the 135.1 image FPS above is measured on A800 with a single prompt. We have stated the GPU, input resolution, and prompt or object workload for the new timings. We have documented the earlier H20 image measurement as multi-prompt because the exact number of prompts and decoding schedule are unavailable. The A800 dedicated video timing and the A800 system benchmark are separate runs. See the video workload (Line 297-301), the single-prompt timing summary (Line 316-318), and the timing table and protocol details (Line 1244-1258).

3. CCA vs. attention on real videos.(Q2)

We added a comparison. The models share a frozen FastSAM-x-initialized encoder, decoder and prompt/memory encoders; attention uses three 256-d, 8-head layers. Training uses 50k SA-V videos, and 1024² resolution.

InteractionSA-V J&FLVOS v2 J&FFPSParams (M)GFLOPs
Global attention75.273.866.180.8261.7
Local attention (w=9)74.672.560.280.8261.7
CCA74.972.994.976.0198.2

Looking at the results, Global attention can achieve highly competitive accuracy (e.g., 75.2 on SA-V) due to its unconstrained receptive field, but at the cost of much lower throughput (66.1 vs. 94.9 FPS) and higher GFLOPs (261.7 vs. 198.2). CCA successfully bridges this gap, achieving a superior quality–speed trade-off: it maintains accuracy on par with or better than local attention while running significantly faster than both attention variants at high resolution. The main-text summary is at Line 407-409, and the attention setup and results are at Line 1266-1283.

AuthorsOpenReview ↗

Response to Reviewer waFd (2/2)

4. EMA and memory-compression sensitivity.(Q3)

Due to time constraints in preparing this rebuttal, we conducted a reduced-budget sensitivity analysis (1024² resolution, 10k iterations, single seed) based on our initial corrected runs. We have reported the measured scores for α=0.9 (23.2 J&F on SA-V val and 37.8 on LVOS v2 val) and α=0 (22.0 and 35.9, respectively) under this protocol. An overly high α=0.99 slightly degrades accuracy. Meanwhile, the FPS remains identical across all α values as it does not alter the forward FLOPs.

For the memory-compression dimension, sweeping values from 128 to 256 shows that all reported paired 95% confidence intervals against the 160-d default contain zero. The FPS remains stable, and the memory bank footprint is unaffected because each entry is expanded back to the 640-channel feature dimension before storage. We have included the reduced-budget sensitivity results and explained that 160-d is the bottleneck width before expansion for storage. The memory path is described at Line 258-265, and the reduced-budget results are at Line 1285-1295.

5. Table 9 and Theorem 1.(Q3)

We have corrected Table 9 using the actual CCA implementation profiler. For FastSAM2 (width 1.0, P4 input [1, 640, 64, 64], CCA dimension 256, k=9, memory bank size 6), the CCA-only costs for L=1/2/3/4 are 13.52/21.68/29.83/37.98 GFLOPs and 1.06/1.63/2.20/2.77 M parameters, respectively. These are per propagated frame at 1024×1024 with batch size 1; the six memory entries are already included, and the 8-frame training clip is not an additional multiplier. Table 6's 0.51G, in contrast, is the analytical estimate for the dominant interaction term O(NTdk²) at N=4096, T=6, d=256, and k=9, rather than an implementation-level count of the complete CCA block. We have made the counting scopes explicit in the captions and accompanying text. In Theorem 1, we have specified a compact feature domain (closed and bounded in finite dimensions) and clarified that the result concerns expressivity of the operator family. See the theorem and its scope (Line 234-238 and Line 651-661) and the two cost tables (Line 714-725 and Line 1047-1060).

We appreciate the reviewer identifying these gaps. We have presented the system and operator results with their respective evaluation settings.

Thank you again for your time and effort in reviewing our submission.

We would like to kindly follow up on our rebuttal, as the rebuttal period is approaching its deadline. We have carefully addressed each of your concerns point by point in our response, and would greatly appreciate it if you could take a moment to review our rebuttal and share any further thoughts or clarifications you may have.

We would be happy to address any remaining concerns or questions.

Thank you very much for your time and consideration!

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

FastSAM2 shares a CNN encoder and mask decoder across images and videos, using CCA for video memory interaction. CCA combines temporal mixing, similarity-gated spatial aggregation, memory weighting and fusion. Tables 2–3 report about 3.2 times the video FPS of SAM 2 B+ at similar accuracy on four benchmarks, with lower accuracy on LVOS v2.

Critical strengths
  1. Tables 2–3 report a useful speedup over SAM 2 B+ with comparable accuracy on four video benchmarks, measured across three GPUs.
  2. The experiments include component ablations and long-video evaluation. Table 5 supports the temporal convolution and fusion gate; Table 3 also makes the long-video accuracy cost clear.
Critical weaknesses
  1. Limited methodological novelty. FastSAM established the CNN-based route to image Segment Anything, while DECO (ICLR 2025, Secs. 3.1 and 3.3) already used convolutional interaction in SAM. The authors acknowledge this lineage. CCA extends it to frame–memory interaction through temporal mixing, similarity gating, memory weighting and fusion. The submission describes these additions, but does not establish a substantial methodological advance over this prior work. Its strongest contribution is the system-level speedup; the novelty case for CCA is weak.
  2. The operator definition and ablation setup need clarification. Eq. (2) gates V_q with Q_p and K_q, whereas Eq. (9) gates it with Q_q and K_q before convolution. It is unclear which version was used in the experiments. Table 9 lists 34.5M parameters at L=3, but the default model has 76M and lite is defined with L=2. Appendix L also uses eight-frame, SA-V-only training. The paper needs to identify the ablation model and explain how these results relate to the main jointly trained model.
Questions for authors
  1. DECO already applies InterConv to SAM. What is the main new mechanism in CCA? Please separate the changes needed to handle video memory from the changes you regard as a new interaction design.
  2. Was CCA compared with an InterConv-based memory module while keeping the backbone, decoder and training fixed? If so, please give the SA-V and LVOS v2 scores and FPS, and explain how InterConv was adapted. If not, please say so.
  3. Table 9 lists 34.5M parameters at L=3, but the default model has 76M and lite is defined with L=2. Which model was used for these ablations? Also, Appendix L uses eight-frame, SA-V-only training. Were Tables 4–5 and 8–9 all run with that recipe, and were the ablated models retrained for the same number of steps?
  4. Eq. (2) gates V_q with Q_p and K_q, whereas Eq. (9) gates it with Q_q and K_q before convolution. Which version was used for the reported results? A few lines of Stage 2 pseudocode would make this clear.
3 comments
AuthorsOpenReview ↗

Response to Reviewer xPwr (1/2)

We thank the reviewer for the careful assessment and for recognizing the practical speedup and the value of our ablations. We agree that the paper should more clearly distinguish the contribution of CCA from prior convolutional interaction, and that the equations and ablation configurations must be fully consistent with the implementation.

1. Positioning and novelty relative to FastSAM/DECO.(Q1)

We do not claim that convolutional interaction itself is new. Our contribution is to formulate it for recurrent video-memory aggregation, where the operator must combine the current frame with multiple historical memory entries and control noisy temporal evidence. CCA combines (i) temporal depthwise mixing, (ii) similarity-gated local propagation, (iii) learnable weighting across memory entries, and (iv) gated residual fusion. We have revised the related-work and contribution statements to describe CCA's video-memory design and its relationship to FastSAM and DECO (Line 072-077 and Line 125-128).

2. Matched InterConv comparison.(Q2)

We added a comparison that changes only the memory-interaction operator:

OperatorSA-V J&FLVOS v2 J&FFPSParams (M)GFLOPs
InterConv73.770.9113.074.1191.1
CCA74.972.994.976.0198.2

We confirm the SA-V val split is used, evaluation resolution is 1024×1024, and all entries share the same unified evaluator. The encoder is CSPDarknet+PAN-FPN (width 1.0), initialized from FastSAM-x and frozen; the prompt encoder, decoder, memory encoder, and operator are trained from scratch. Both models use L=3, k=9, dimension 256, and six memory entries. InterConv replaces each CCA block with a DECO-style block: averaged memory is added to the current feature, followed by 9×9 depthwise convolution and a 1×1 feed-forward layer with residual connection. It has no temporal mixing, similarity gate, per-entry weighting, or gated fusion. Training uses 10k iterations, batch 4×8 frames, 1024², 50k SA-V videos, same seed. CCA improves J&F by 1.2/2.0 points on SA-V/LVOS, at lower throughput (94.9 vs. 113.0 FPS). The revised manuscript presents the comparison in the ablation section (Line 406-409) and the matched setup and results in Appendix R (Line 1259-1283).

AuthorsOpenReview ↗

Response to Reviewer xPwr (2/2)

3. Table 9 and configuration consistency.(Q3)

The depth ablation in Table 9 was run with the default FastSAM2 configuration (width 1.0, CCA dimension 256, memory bank of 6 entries, k=9), not FastSAM2-lite; only the number of CCA layers L varies, and the encoder, prompt encoder, mask decoder, and memory encoder are identical across rows. The Params/GFLOPs columns of the submitted Table 9 were filled in incorrectly: the L=3 row (34.5 M, 81.7 GFLOPs) coincides with the whole-model figures quoted for FastSAM2-lite and does not describe the ablated models, which caused the inconsistency the reviewer points out. We corrected the table to report CCA-only costs profiled from the actual implementation, using a 1024×1024 input, a 64×64 stride-16 P4 feature map ([1, 640, 64, 64]), CCA dimension 256, k=9, and a memory bank of 6 entries:

CCA layers (L)CCA params (M)CCA GFLOPsSA-V val J&F
11.0613.520.69
21.6321.680.72
32.2029.830.75
42.7737.980.75

These are implementation-profiled CCA-only FLOPs for one propagated video frame (batch size 1). The six memory entries are already included in each measurement, so T=6 must not be multiplied in again. Likewise, the 8-frame training clip length is not an additional multiplier: these are per-frame costs, not clip-level totals. Each additional CCA layer adds approximately 0.57 M parameters and 8.15 GFLOPs; all non-CCA modules are unchanged and excluded from these columns.

We also clarify the apparent difference from Table 6. Table 6's 0.51G is an analytical estimate of the dominant local-interaction term obtained from O(NTdk²), with N=4096, T=6, d=256, and k=9. It is not an implementation-level count of the complete CCA block. By contrast, Table 9 profiles the actual CCA implementation, including the input/output projections and other implemented stages (the P4 input has 640 channels before projection to the 256-dimensional CCA space). Thus, the two tables report different scopes and counting conventions, rather than conflicting measurements of the same quantity. We have revised the Table 6 caption and text to identify 0.51G as an analytical multiply-accumulate estimate for the dominant interaction term. We have labeled the Table 9 values as implementation-profiled CCA-only GFLOPs per propagated frame. The analytical count is explained in Appendix B.2 (Line 714-725), and the corrected profiling values and scope are in Appendix L (Line 1024-1025 and Line 1047-1060).

Regarding the training recipe: Tables 4–5 and 8–9 all use the Appendix L recipe (8-frame clips, 1024×1024, SA-V train only), and every ablated model was retrained from the same initialization for the same number of 50 epochs. We have stated the configuration, counting convention, and recipe in the captions and accompanying text. We have also corrected the conflicting descriptions of FastSAM2 and FastSAM2-lite (Line 746-755). The ablation recipe and captions are in Line 378-379, Line 1015-1019, and Line 1047-1050.

We have scoped the claim to the measured efficiency–quality trade-off of FastSAM2 and the specific design choices evaluated (Line 406-409 and Line 421-425).

4. Equations (2) and (9).(Q4)

We checked CCA in the released implementation. The gate is computed between Q and K at the same spatial position, channel-wise; neighborhood aggregation occurs only in the subsequent k×k depthwise convolution. We have corrected Eq. (2) to reflect this operation (Line 198-215).

  • Gate (same position, per channel):G(i)(q)=σ!(Q(q)⊙K(i)(q))∈(0,1)C.G^{(i)}(q)=\sigma\\!\left(Q(q)\odot K^{(i)}(q)\right)\in(0,1)^C.
  • Local aggregation:A(i)(p)=∑q∈Nk(p)Wdw(p−q)⊙G(i)(q)⊙V(i)(q).A^{(i)}(p)=\sum_{q\in\mathcal{N}_k(p)}W^{\mathrm{dw}}(p-q)\odot G^{(i)}(q)\odot V^{(i)}(q).
  • Output:Corri=SE!(SiLU!(BN!(A(i)))).\mathrm{Corr}_i=\mathrm{SE}\\!\left(\mathrm{SiLU}\\!\left(\mathrm{BN}\\!\left(A^{(i)}\right)\right)\right).

We have added BN and SiLU after the depthwise convolution in the appendix equation and included implementation-matched pseudocode. The depthwise kernel depends on offset and channel. The gate provides content dependence at position q. See the corrected appendix equation (Line 766-770) and the pseudocode (Line 783-790).

# f3d: [B,C,H,W] Stage-1 output; m_i: [B,C,H,W] i-th memory entry
q = SiLU(BN(Conv1x1(f3d)))       # [B,C,H,W]
k = Conv1x1_K(m_i)               # [B,C,H,W]
v = Conv1x1_V(m_i)               # [B,C,H,W]
gate = sigmoid(q * k)            # same-position, channel-wise product (no neighbor search)
x = DWConv_kxk(gate * v)         # groups=C, k=9, padding=4: local aggregation over N_k(p)
corr_i = SE(SiLU(BN(x)))         # [B,C,H,W]

We have revised the manuscript and narrowed our claim to the measured system-level trade-off and the specific CCA design choices evaluated.

Thank you again for your time and effort in reviewing our submission.

We would like to kindly follow up on our rebuttal, as the rebuttal period is approaching its deadline. We have carefully addressed each of your concerns point by point in our response, and would greatly appreciate it if you could take a moment to review our rebuttal and share any further thoughts or clarifications you may have.

We would be happy to address any remaining concerns or questions.

Thank you very much for your time and consideration!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.