- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes a temporal latent representation for long binary quanta image sensor (QIS) sequences. Instead of reconstructing frames at fixed timestamps prior to compression, a hierarchical encoder aggregates binary photon streams into multi-scale temporal latent anchors. At the receiver side, a temporal query module interpolates neighboring anchors with learned residual refinement to decode frames at arbitrary timestamps, while cross-scale predictive coding compresses the latent state for transmission. The method is evaluated across five simulated photon levels and on real SPAD sensor captures, alongside analyses of payload size and encoding/query runtime.
Critical strengths
-
Encoding the binary measurement stream once into a queryable latent state is a very sensible design for QIS transmission. Decoupling sender-side photon aggregation from receiver-side timestamp selection avoids committing to a fixed frame rate upfront and fits well with the high-speed nature of binary photon streams.
-
The evaluation covers five simulated photon-rate settings, α ∈ {0.05, 0.1, 0.2, 0.3, 0.5}, as well as real QIS recordings. I appreciate the reported serialized latent payload of 4.628 MiB for a 4096-frame sequence and the separation of sender encoding, receiver initialization, and per-query costs, including the explicit qualification of Bit2Bit’s sample-specific optimization.
Critical weaknesses
-
Key details regarding how the baselines were adapted and evaluated are missing, which makes the ~3 dB PSNR gain over QUIVEr hard to verify. In the original QUIVEr paper, simulated experiments primarily use 3-bit inputs under a different noise setup, whereas here the input is 1-bit Bernoulli data with hot-pixel and dark-count defects. It is unclear whether QUIVEr and other baselines were retrained from scratch on the new 200-video training set or tested using pretrained weights, and how their temporal windows, intensity scaling, and defect handling were adjusted. Also, since the 200/13 dataset split differs from QUIVEr’s original 249/31 split, we need to be sure that none of the 13 test videos were seen during baseline pretraining. Finally, it is not fully clear whether "averaging over all reconstructed frames" evaluates all methods at the exact same target timestamps and effective temporal windows.
-
The advantage of querying in latent space is not cleanly separated from simply having a stronger base reconstructor with a longer temporal receptive field. A very natural baseline is missing: running a strong QIS reconstruction method at sparse keyframes, compressing them, and applying an off-the-shelf video frame interpolator at the receiver for arbitrary timestamps. Without comparing against "sparse reconstruction + frame interpolation" at matched bitrates and query times, it is hard to tell how much gain actually comes from the continuous latent query mechanism itself.
-
Appendix F explicitly states that the timing examples compare workflows with different native output counts and include sample-specific optimization for Bit2Bit; these results should therefore be interpreted within that scope. The submitted source package also provides substantial architecture and training configurations. The remaining reproducibility issue is their mapping to the reported experiments: for example, the reconstruction configuration uses batch size 4, a warmup/plateau schedule, and loss weights that differ from Section 4.1, while the README describes independently trained reconstruction and compressed variants initialized through a linear-anchor run. Please identify the exact configuration and checkpoint lineage for each reported model and reconcile these differences with the manuscript.
Questions for authors
-
Were the baseline methods (especially QUIVEr) retrained on your 1-bit 200-video split, or evaluated using released checkpoints? How were bit-depth differences, photon levels, intensity scaling, and sensor defects handled for each baseline?
-
If any pretrained weights were used for the baselines, can you confirm there is no overlap between their original training sets and your 13 test videos (e.g., by listing the test video IDs)?
-
In Table 1, are all methods evaluated at the exact same timestamps and ground-truth frames, and how does this align with the different native output frame counts noted in the runtime comparison?
-
How would your method compare against sparse reconstruction followed by video frame interpolation under matched payloads and query timestamps? Please also identify which submitted configuration produced each reported model, including its initialization lineage, interpolation variant, and final 13-video evaluation procedure.
Response to Author
The authors have satisfactorily addressed all of my concerns. Therefore, I will maintain my original score.