acceptodds
Under review as a conference paper at ICLR 2027

Event Generative Compression

Abstract

Event cameras offer high temporal resolution and dynamic range. However, in dynamic scenes, they can generate event streams with data rates more than two orders of magnitude higher than conventional video, making transmission a major bottleneck for remote perception. While existing approaches primarily aim to preserve the event stream itself, we argue that when the downstream consumer only needs visual content, compression should instead preserve information useful for visual reconstruction and perception, rather than events themselves. We propose Event Generative Compression (EGC), which encodes each temporal window into a compact set of discrete symbols and uses a pretrained video diffusion model to recover the missing visual information. EGC learns an event-to-generation interface through latent alignment, discrete quantization, entropy coding, and generative adaptation, while reusing generated history for continuous reconstruction without transmitting reference frames. On DSEC and BS-ERGB, EGC achieves compression ratios of approximately 5,000–8,000 relative to losslessly coded event streams while retaining information for both temporally coherent video reconstruction and downstream tasks. The same low-rate messages support semantic segmentation, optical flow, and depth estimation, demonstrating the effectiveness of EGC in preserving useful visual information under ultra-low bitrate compression.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Event Camera, Video Compression

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_79zbOpenReview ↗
Rating
4
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes Event Generative Compression (EGC), a cross-modal event compression framework in which an event-camera sender transmits a highly compressed discrete latent message and a receiver uses an adapted pretrained video diffusion model to reconstruct RGB video. The method combines event-latent alignment, finite scalar quantization with a 3D checkerboard entropy model, residual flow matching, and generated temporal history, and is evaluated for reconstruction, segmentation, optical flow, and depth on DSEC and BS-ERGB.

Critical strengths

The paper poses a useful alternative to conventional event coding: preserve the information needed for visual reconstruction and downstream tasks rather than requiring event-stream fidelity. This is a meaningful formulation for settings where the receiver consumes visual content rather than events

The system design is coherent. Figure 2 clearly separates the event-only sender, exact entropy-coded symbol transmission, and the receiver-side generative prior. In particular, the architecture makes the operational assumption clear: no RGB frame or transmitted reference frame is available to the sender at inference. Figure 3 also provides a helpful account of what is optimized and frozen at each of the three training stages.

The temporal-history mechanism is supported by both visuals and metrics. visible window-boundary color discontinuity without history that is mitigated by generated history, while Table reports substantial decreases in the seam measure on both DSEC and BS-ERGB. This is a well-chosen ablation because temporal continuity is a central practical challenge once no reference frames are transmitted.

The empirical rate regime is striking.Table 3 verifies that the learned checkerboard context model actually reduces the measured container rate relative to zstd while preserving exactly the same symbols. This is much stronger than reporting only estimated entropy or nominal FSQ capacity.

Critical weaknesses

The baseline suite is incomplete for the paper's central claim about event compression. The comparison emphasizes reconstruction methods combined with random retention, temporal downsampling, or GLC, but does not evaluate against a dedicated event-stream compression method such as flow-based visual stream compression. Rezaee et al. is cited as lossy event compression work but is not included experimentally. This is important because EGC’s stated advantage is not only good event-to-video reconstruction, but substantially better preservation of useful visual/task information at a given event-transmission budget. A direct event-compression baseline should be encoded at matched rates and evaluated through the same reconstruction and downstream pipelines. Otherwise, it remains unclear how much of the gain derives from a stronger receiver prior versus the proposed learned compressed message.

Stage 3 trains with zero or perturbed ground-truth history, whereas inference conditions on the model's own increasingly imperfect output. This can produce long-horizon drift, identity swaps, or temporal inconsistency not captured by two boundary examples or a seam statistic. The paper should report metrics over substantially longer rollouts, error as a function of window index, and recovery behavior after a reset or corrupted history. This is particularly important because the method is presented for continuous remote perception.

This may be acceptable for an offline or server-side receiver, but it is a substantial limitation for low-latency remote perception. The paper should report receiver memory, end-to-end latency including coding/transport/decoding, throughput at the stated resolution, and at least a discussion of whether a less expensive generative receiver changes the rate-quality conclusions.

Questions for authors

Please provide the complete Stage-1, Stage-2, and Stage-3 objectives in the main paper, including all coefficients, the FSQ parameterization, the flow-matching schedule the time sampling distribution, and the exact treatment of history during training.

Can the authors compare to a dedicated event-stream compression baseline at matched rates, using the identical reconstruction receiver and downstream task protocols? This would directly test the paper’s compression claim.

How much do reconstruction and task metrics vary across diffusion sampling seeds? history handled across adjacent windows ?, and what happens after long rollouts, a receiver restart, or a corrupted/generated-history mismatch?

2 comments
AuthorsOpenReview ↗

Long rollouts, history recovery, receiver cost, and additional comparisons

Thank you for your assessment and suggestions. We have clarified the rollout length and added history-intervention results and receiver timing measurements. The revised manuscript includes the additional reconstruction and downstream comparisons in Figures 8 and 9.

1. Long rollouts and history feedback. The 121-frame length refers to one generation window. Our DSEC full-stream evaluation includes a continuous rollout of 25 windows, producing 2,827 frames over 113.08 seconds. History is updated from the model's own outputs throughout this rollout. Section A.3 now states the sequence length explicitly.

To examine error propagation, we evaluate a one-time history perturbation and a history reset at the same selected window of this DSEC sequence. After the intervention, normal generation resumes and the affected outputs continue to supply history to later windows. The comparisons use the same event inputs and paired sampling settings. We report reconstruction quality over time and the difference from the unperturbed rollout to distinguish the immediate effect from its persistence in later windows. Section D.3 and Figures D1–D2 now report these results. Normal-rollout quality fluctuates without sustained deterioration over this sequence. At window 9, Gaussian noise with standard deviation 0.5 is added to the most recent history position, or all eight history positions are reset to zero. With noise, LPIPS distance to the paired normal rollout decreases from 0.00990 at window 9 to 0.00231 at window 11, then rises to 0.00851 at window 25. Reset causes a larger difference, decreasing from 0.14943 at window 9 to 0.03860 at window 25; the reset branch does not fully return to the normal trajectory.

2. Receiver cost and first output. Section F.3 adds a measurement from events in host memory through actual CCM bitstream construction to reconstructed RGB in host memory. On one GPU at 640 × 464 resolution with 25 sampling steps, an independent 121-frame window takes 4.015 s at the sender and 17.880 s at the receiver, totaling 21.895 s of processing. Collecting the first event window takes 4.84 s. Assuming a net transmission rate of 10 Mbit/s, its 14,537-byte bitstream takes approximately 0.012 s to transmit. The first reconstructed window is therefore available after approximately 26.75 s in this sequential setting. Processing time is measured; transmission time is estimated from the stated bandwidth. Model loading and warm-up are excluded.

The receiver throughput is 6.77 fps for this independent window and 6.22 fps for the eight-window measurement. The separately measured receiver peak allocated GPU memory reported in Section F.2 is 20.133 GiB. The current receiver requires buffering and does not keep pace with a continuous 25 fps input on this GPU. The intended use is bandwidth-limited transmission where receiver-side computation and delayed reconstruction are acceptable. Moving video generation to the receiver reduces computation at the event-only sender (Table 1a). A faster or smaller receiver is an important next step; its rate–quality trade-off needs a separate evaluation. Section F.3 also reports throughput and first-output measurements for other diffusion-based reconstruction methods, with their measurement settings specified.

3. Additional compression comparisons. We have completed reconstruction and downstream comparisons for JPEG2000, G-PCC, x265/HEVC, and DCVC-FM. These cover event-domain compression followed by reconstruction and reconstruction followed by video compression. Main-text Figure 8 reports grayscale LPIPS against measured bitrate, and Figure 9 reports the three downstream tasks with shared frozen task networks for the RGB routes. Supplementary Section C.7 gives codec settings and the corresponding Q comparison. The fixed-receiver comparison in main-text Table 2b and Supplementary Table C3 continues to isolate the transmitted representation using the same generative receiver.

4. Training details and sampling seeds. We have added the straight-through treatment of quantization, the flow-matching time-sampling distribution, the 25-step Euler schedule with shift 5, and the velocity-loss definition to main-text Sections 3.2–3.4. Supplementary Section A.2 gives the complete objectives and coefficients. History is zeroed with probability 0.5 during training; otherwise Gaussian noise with standard deviation 0.1 is added to ground-truth history. Inference uses generated history. Supplementary Figure D2 also reports an additional seed through window 14, showing a similar initial decrease in the effect of history noise.

AuthorsOpenReview ↗

Thank you for your feedback and reassessment

Thank you for taking another look at the paper and for revising your assessment. We have included the additional results and training details in the revision, and made their locations easier to find in our response. We appreciate the time you have spent discussing the work with us.

Reviewer_b3XuOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Event Generative Compression (EGC), which compresses events for transmission. The sender compresses event representations into compact latents using finite scalar quantization and entropy coding, while the receiver dequantizes them and generates RGB videos with Wan2.2-5B conditioned on the event latents. On DSEC, the compression ratio can reach approximately 8,000:1 by comparing the 0.004283 bpp bitrate of the transmitted latents with the reported 35.30 bpp full-event rate after general-purpose lossless coding. After compression, the videos can still be reconstructed with considerable quality.

Critical strengths
  • The research topic is interesting and promising. For example, in long-distance communication scenarios such as space communication, the remote edge device and the ground station may both have considerable computational resources, while communication bandwidth remains the bottleneck. In such settings, a compression method of this kind could be very important.

  • The reported compression ratio of approximately 8,000:1 looks impressive.

Critical weaknesses
  • The most important issue is that all experiments and metrics concerning information preservation appear to be conducted on reconstructed frames. Line 268 states that “the latter operates directly on generated latents without RGB decoding,” but the paper does not explain which tasks use this protocol or how these experiments are conducted. However, the event stream contains a large amount of information that is not present in conventional RGB video, such as information at a much higher temporal resolution. If the method is evaluated only on videos with a particular frame rate—which is not specified in the paper—the results cannot demonstrate that information beyond the RGB video has been preserved. According to Line 51, H.264-encoded RGB video requires only 2–4 Mbps. Assuming that this refers to the complete DSEC RGB stream at 20 FPS, this corresponds to approximately 0.32 bpp. Relative to the RGB video bitrate, EGC therefore achieves only roughly a 1/80 compression ratio.

  • Although the method uses strong priors from a video generative model to supplement the compressed information, and although DSEC and BS-ERGB cover relatively constrained domains dominated by road-driving scenes, leaving room for substantial compression, the final decoded videos still appear to contain many important details that are less accurate than those produced by conventional reconstruction methods, according to the qualitative results. This dataset bias may also make the learned representation less suitable for other domains. In particular, events do not contain color information, so the colors in the reconstructed RGB videos are largely hallucinated. This makes me question the value of generating RGB videos in this setting.

  • The following comments do not affect my overall assessment and are only provided as suggestions for revision:

    • Some inconsistent or unclear notation causes confusion:
      • Lines 165–168: T′T', H′H', and W′W' are not defined. Are they intended to satisfy T′=TT'=T, H′=HH'=H, and W′=WW'=W?
      • Lines 203–208: Does the symbol zcz_c have the same meaning as z^rgb\hat{z}_{\mathrm{rgb}}?
    • Several abbreviations are not explained when they first appear:
      • bpp: bits per pixel
      • tOF: temporal optical flow
      • tLP: temporal learned perceptual similarity
    • These issues make the paper more difficult to read.
Questions for authors
  • Regarding the statement, “The latter operates directly on generated latents without RGB decoding,” could the authors provide more details about these experiments?
4 comments

I thank the authors for the supplemented information, but would still like to discuss my concerns.

  1. Information loss of the representation

“by shortening event bins from 40 to 20 and 10 ms”: Does reconstructing at higher frame rates require re-encoding the event streams, and using higher bit-per-second rates? Does the "bin" refer to the "interval" from "defined over the T output-frame intervals of the window"(L165)?

Please correct my understanding if I am wrong: at the beginning of the encoding, information about events within each bin * each pixel is crushed into a 2 bit number in {-1, 0, 1} (all intermediate time resolution information is lost); Thus, after encoding and decoding, for the time interval corresponding to each original bin, only one single RGB frame or depth / segmentation / optical flow frame can be predicted.

  1. Comparison against HEVC.

I now understand that although video compression can achieve lower bitrates, but it corresponds to much lower reconstruction quality. If I were to conduct the experiments myself, I would probably choose E2VID+ reconstructions to be push the compression rates further down, since EvDiff, as a diffusion model, hallucinates many high-frequency details that are hard to compress. And I would provide a bpp-CRF-LPIPS curve to show the trade-off of quality and bitrate as compression strength goes up. However I think the current results as good enough.

  1. Generalization.

Another of my critical concerns is whether the encoding model, trained on driving scenes of DSEC and BS-ERGB, can generalize to other scenes (e.g. the classic indoor office scenes of MVSEC). Seeing such qualitative examples would be convincing for me.


Tips for presentation: to make understanding easier for human readers, you may highlight best results in tables, and explain terms (such as "Route A/B/C refer to experiments that can be found in the supplementary material") when they appear. Also, you can post multiple comments if your content exceeds 5000 chars.

AuthorsOpenReview ↗

Temporal resolution, compression comparisons, and indoor results

Thank you for following up. We should have explained the temporal bin more clearly in our earlier reply. We have also added the compression comparisons and the indoor example you suggested. We have updated both the main paper and the supplementary material.

1. Temporal information in the representation.

Yes, each bin corresponds to one output-frame interval. At each pixel, we keep the polarity of the last event in that interval, or zero if there was no event. This discards the individual timestamps and event counts within the bin, so an encoded stream fixes the temporal grid for reconstruction and the associated task outputs.

The 50 and 100 fps examples in Supplementary Section C.6 require re-binning and re-encoding the original events at 20 and 10 ms. They cannot be obtained from the existing 25 fps bitstream. With the same spatial and token configuration, a finer grid means encoding more tokens per second, so the nominal bitrate in bits per second increases.

What we meant by flexibility is that the sender can choose the output interval from the event stream before compression, according to the application's temporal and bandwidth requirements. We agree that encoding once and allowing the receiver to choose the reconstruction frame rate would be more useful. We would like to explore this in future work.

2. Reconstruction followed by video compression.

We have added E2VID+, ET-Net, and EvDiff followed by x265 to main-text Figure 8, which plots grayscale LPIPS against measured bitrate. These x265 settings vary spatial resolution and frame sampling at a fixed CRF51; the caption and Supplementary Section C.7 specify the settings.

For the indoor example in Figure 10, we compare EGC with E2VID+ followed by HEVC at a similar bitrate (0.0075 vs. 0.0077 bpp for EGC), and also include a lightly compressed HEVC version (CRF23).

3. Generalization to indoor scenes.

Following your suggestion, we added a qualitative comparison on MVSEC indoor_flying1 in Figure 10. MVSEC was not used to train EGC. The figure shows eight matched time points from 0.51 to 5.86 s relative to the segment start, approximately 0.765 s apart, with ground truth, the frozen EGC model, and the two E2VID+→HEVC settings above. EGC uses interframe events and per-frame Gray World correction without fitting brightness to GT. At the nearby measured rates, EGC retains more visible scene structure than the heavily compressed E2VID+ video in these frames.

The supplementary video now includes this MVSEC comparison immediately after the earlier higher-frame-rate demonstration.

We have also explained Routes A/B/C where they first appear in the main text and highlighted the best and second-best values within comparable table groups, as you suggested.

Thank you for the detailed clarification. My concerns have been resolved, and I have updated my rating accordingly.

Thank you for the time you put into reading and discussing our paper. Your questions about temporal resolution and generalization gave us a lot to think about, and your suggestions on the writing helped us explain the work more clearly. We really appreciate your patience and the helpful exchange.

Reviewer_mt4bOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper looks at extremely low-bitrate compression for event cameras and treats it as a cross-modal problem: the sender only sees events, and the receiver wants RGB video and task outputs rather than the event stream itself. In the proposed EGC framework, each window of events is turned into a small quantized, entropy-coded latent message. On the receiver side, a pretrained video diffusion model (lightly fine-tuned) generates RGB video from this message and from its own previously generated frames, so no reference frames are sent. Training is done in stages: aligning events to the video latent space, then quantization and entropy coding, then adapting the generator. On DSEC and BS-ERGB the reported rates are several orders of magnitude below lossless event coding, with perceptual quality that the paper describes as competitive with or better than reconstruct-then-compress cascades. There are also downstream results (segmentation, optical flow, depth), component ablations and compute numbers.

Critical strengths
  1. A novel problem formulation, supported by a working end-to-end system. A message of a few tens of kbps is enough to drive a large video prior (22.37 kbps full-stream for the base model on DSEC, Table C2). With the receiver fixed, the learned message matches the LPIPS of ESR+x265 at 0.003 bpp versus 0.86 bpp (Table 2b / C3). On DSEC, EGC beats every evaluated cascade at a lower rate (Fig. 4).

  2. Rigorous rate accounting. The rates are measured bytes including metadata and overlap, all 393 containers decode exactly, and the authors also report the two configurations where the context model does not help. The wrong-message, mean-latent and zero-index interventions confirm that the output depends on the transmitted content: on DSEC, LPIPS goes from 0.51 to between 0.62 and 0.86 (Table D1).

Critical weaknesses
  1. The baselines do not establish how EGC compares with alternatives near its own operating rates.

    • There is only one reconstruct-then-compress baseline (GLC). It is evaluated only around 10−210^{-2} bpp, and its rates are likelihood estimates rather than measured bytes. On BS-ERGB, EGC's best point (Q = 62.5) is well below E2VID+→GLC and LongE2V→GLC (roughly 80–92 in Fig. 4), which the paper acknowledges. Results for GLC or a standard video codec at lower rates, or an event-domain lossy codec, with overlapping rate ranges are needed to resolve this frontier.
    • EGC is trained on BDD100K, DSEC and BS-ERGB with RGB targets and also gets Gray World correction. The paper does not say whether the E2V baselines were trained on the same data.
    • Q = 93.2 means "best among the evaluated points", not high fidelity. In absolute terms, the headline configuration reaches 13.87 dB PSNR / 0.468 LPIPS on DSEC (Table C2).
    • The optical flow advantage comes from Route A, where task heads are trained with DSEC labels, supervision unavailable to the baselines. The like-for-like comparison (Route C) is roughly on par with the GLC baselines and is only shown in a plot.
  2. The practical premise is only partly supported.

    • Windows are 121 frames (about 4.8 s), and the receiver runs at about 7 fps on an H200. This is hard to reconcile with the remote-perception motivation, and end-to-end latency is not reported.
    • Quantization incurs a substantial loss of task-relevant information. Compared with continuous conditions, mIoU drops from 42.62 to 37.26 and flow EPE rises from 1.556 to 2.370 (Table 1c).
    • Error propagation is not studied: no experiment feeds a perturbed window back into the history. Quality is also not monotonic across configurations (e.g., 2222-ts1 gets Q = 89.4 at 0.0022 bpp vs. 71.6 for base at 0.0030 bpp, Table C1), which weakens the "more capacity helps" claim in Sec. 4.3.
Questions for authors

see the weakness part.

3 comments
AuthorsOpenReview ↗

Responses to the comparison and practical-use concerns

Thank you for your comments. We have completed the reconstruction and three-task comparisons for all four additional codecs and added them to main-text Figures 8 and 9. Both the main paper and supplementary material have been updated; we detail the changes below.

W1: How does EGC compare with baselines near its operating rates?

Our GLC comparison includes q=0, the lowest rate setting of the released model. We added four baselines to examine rates near EGC.

JPEG2000 compresses event histograms, and G-PCC compresses event point clouds, both before reconstruction. x265 (HEVC) and DCVC-FM (learned video compression) instead compress reconstructed video. Figure 8 reports grayscale LPIPS against measured bitrate on DSEC and BS-ERGB; Supplementary Section C.7 gives the settings and the corresponding Q comparison.

Representative results use EvDiff throughout. JPEG2000/G-PCC use their lowest tested setting; x265 uses half-resolution CRF51. Each cell reports bpp / gray LPIPS:

MethodDSECBS-ERGB
JPEG2000→EvDiff0.005343 / 0.75520.004036 / 0.7162
G-PCC→EvDiff0.003123 / 0.75530.001070 / 0.8356
EvDiff→x2650.002703 / 0.75070.002940 / 0.9006
EvDiff→DCVC-FM q00.002199 / 0.69540.003132 / 0.8551
EGC base0.003033 / 0.50990.003247 / 0.6168

EGC has lower LPIPS in these comparisons.

W1: Were the baselines trained on the same data, and does Gray World affect the comparison?

The original baselines use released weights. We trained E2VID+ from scratch on EGC's data sources: BDD100K, DSEC, and BS-ERGB, using 22,294 clips of 20 frames and 11,148 updates. Both E2VID+ versions use uncompressed events. Supplementary Table C6 (Section C.8) gives all five reconstruction metrics; below we show gray LPIPS / MS-SSIM:

MethodDSECBS-ERGB
Released E2VID+0.5129 / 0.40590.4370 / 0.5725
Same-source, 11,147 clips0.5590 / 0.36800.5167 / 0.5029
Same-source, 22,294 clips0.5448 / 0.36610.5018 / 0.5332
EGC base0.5099 / 0.55660.6168 / 0.4061

Longer training improves LPIPS on both datasets, but released weights remain stronger on both listed metrics, supporting their use as our baseline at this training budget. EGC uses fixed, GT-free Gray World correction (Section B.1); all downstream evaluations omit it.

W1: Does Q = 93.2 imply high absolute fidelity?

No. Q summarizes relative quality using fixed mappings of grayscale LPIPS, MS-SSIM, tOF, and tLP (Section B.1). The main text now states this explicitly, and Figure 8 reports LPIPS directly.

W1: Does the optical-flow advantage depend on task-head supervision?

Figure 9 includes segmentation, flow, and depth for all four codecs. Their RGB outputs and EGC Route C use the same frozen task networks, with rates measured on each task's sequences (Section E.1). For flow, all use frozen SEA-RAFT. Representative low-rate results are:

MethodMeasured bppEPE ↓
JPEG2000→EvDiff0.00533411.085
G-PCC→EvDiff0.00326310.977
ET-Net→x2650.0035168.003
EGC base, Route C0.0032766.240

These Route C results use no EGC-specific task-head training. The main text now explains Routes A/B/C and their supervision differences.

W2: What is the end-to-end latency?

On one GPU at 640 × 464 with 25 sampling steps, a 121-frame window takes 4.015 s at the sender and 17.880 s at the receiver. With 4.84 s acquisition and estimated 0.012 s transmission at 10 Mbit/s, first-window latency is 26.75 s, excluding loading and warm-up (Section F.3). Continuous input requires buffering; our intended setting allows delayed reconstruction under limited bandwidth. Receiver acceleration remains future work.

W2: Does quantization lose task information?

Yes: relative to continuous conditioning, mIoU drops from 42.62 to 37.26 and flow EPE rises from 1.556 to 2.370 (Table 1c). The shared-network results above show useful task performance at low transmission rates despite this cost.

W2: What about history errors and non-monotonic capacity trends?

The added 25-window rollout generates 2,827 frames over 113.08 s. At window 9, we add Gaussian noise (standard deviation 0.5) to the latest history position or zero all eight positions, then continue with each branch's generated history. LPIPS distance to the paired normal rollout at windows 9/11/25 is 0.00990/0.00231/0.00851 for noise and 0.14943/0.05943/0.03860 for reset. Effects initially diminish but persist. Normal quality fluctuates without sustained deterioration in this sequence (Section D.3, Figures D1–D2). Section 4.3 now describes the capacity trend as an overall improvement with intermediate fluctuations.

AuthorsOpenReview ↗

Updated experiments and revised manuscript

Thank you again for your feedback. We have completed the additional experiments and incorporated them into the revised paper and supplementary material. We also updated our earlier reply with the results.

Figures 8 and 9 now show reconstruction and downstream comparisons with JPEG2000, G-PCC, x265, and DCVC-FM, including RGB-route comparisons using the same frozen task networks. The supplementary material includes the same-source E2VID+ comparison, history-perturbation results, and end-to-end timing measurements.

If you have a chance to look at the updates, we would appreciate hearing whether they address your concerns.

Thank you for the detailed clarification. My concerns have been resolved, I therefore maintain my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.