{"id":"6e0dff3a-5538-46e3-a119-0f3dfa3fe960","arxiv_id":"2607.07823","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.5,"correctness_risk":"low","formal_verification":"none","parameter_count":4,"one_line_summary":"Spatial-domain neural JSCC recovers localized video block erasures via inpainting while feature-domain JSCC recovers distributed erasures via cross-channel semantic redundancy, with measured PSNR trade-offs under uniform and importance-weighted losses.","lead":"Neural video codecs can be trained to survive block erasures by splitting frames into spatial patches or feature groups and learning to fill the gaps. The comparison shows when each split works better, guiding robust real-time video links without retransmissions.","discovery_kind":"extension","skeptic_critique":{"model":"grok-4.5","headline":"Strongest claim rests on non-rate-matched PSNR curves that mix training regimes and bitrates, so the spatial-vs-feature superiority comparison is not cleanly isolated.","rationale":"The reader correctly flags the hand-crafted central-face importance map as a limitation, but that map only affects the two-level non-uniform experiments of §III.B. The strongest claim in the abstract and conclusion is the broader spatial-versus-feature complementarity under both uniform and non-uniform erasures. That claim is supported by the reported curves, yet those curves are confounded by the unmatched rates the paper itself acknowledges. Rate-matching is therefore a more load-bearing experimental control than the semantic map: if the ranking of the two domains disappears once rates are equalized, the headline insight does not hold. The paper remains a solid empirical systems contribution; the concern simply tightens the already-CONDITIONAL recommendation by making rate-matched ablations the primary missing check rather than (or in addition to) broader datasets. No evidence of internal inconsistency or fabrication; the limitation is methodological and openly noted by the authors.","tokens_in":8557,"tokens_out":606,"duration_ms":7145,"concrete_test":"Retrain both the spatial-domain (K=16) and feature-domain (K=8) models under identical target BPP (e.g., by adjusting λ or adding an explicit rate constraint so that clean-channel BPP differs by <0.01) and re-plot the mismatch curves of Figs. 2 and 5; if the relative ranking of the two domains reverses or the gap shrinks by more than ~1–2 dB at the critical low-to-moderate erasure rates, the strongest claim weakens.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central experimental claim (abstract + §§III–IV) that spatial-domain JSCC “excels at handling random localized losses” while feature-domain JSCC “provides superior robustness to distributed erasures and maintains fidelity under low-loss scenarios” is drawn from mismatch curves (Figs. 2 and 5) and non-uniform results whose operating points are explicitly not rate-matched. The paper itself notes “the curves are not rate-matched, with each model settling at a slightly different bitrate” (Fig. 2 caption and surrounding text) and again for the feature-domain models (Fig. 5: BPP values range 0.2959–0.3558). Because rate–distortion trade-offs are controlled by a single λ=2048 and the entropy model is free to adjust, differences in clean-channel BPP and in the amount of learned redundancy can themselves produce the observed PSNR gaps. Consequently the claimed complementarity of the two partitioning strategies is not isolated from rate effects; the superiority statements may partly reflect unequal bit budgets rather than intrinsic spatial-continuity versus semantic-redundancy advantages.","agreement_with_reader":"partial"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes a semantic-aware neural JSCC framework for video transmission over multi-level block erasure channels, built on the DVC architecture. It studies two complementary partitioning strategies: spatial-domain partitioning of motion/residual tensors into 16 localized blocks (with uniform and two-level semantic-guided non-uniform erasures that protect central face regions more than periphery) and feature-domain partitioning into 8 channel groups (with progressive non-uniform erasures). Training injects erasures via a differentiable multi-level interface that replaces erased blocks with a -1 placeholder. Experiments on Vimeo-90k training and VCD talking-head evaluation report PSNR under mismatch (Figs. 2, 5), visual recovery (Figs. 3, 8), regional PSNR (Fig. 4), progressive-block decoding (Figs. 6–7), and conclude that spatial-domain JSCC handles random localized losses better while feature-domain JSCC is more robust to distributed erasures and low-loss regimes.","tokens_in":8899,"tokens_out":936,"duration_ms":8470,"significance":"If the spatial-versus-feature complementarity holds after rate equalization, the work supplies a concrete design insight for task-aware video JSCC over erasure channels that is directly relevant to low-latency conferencing and teleoperation. Strengths include the explicit multi-level erasure interface, systematic mismatch analyses, visual recovery sequences that demonstrate temporal error propagation control, and the progressive-block decoding curves that illustrate unequal error protection under feature-domain non-uniform training. These elements go beyond pure rate–distortion papers and give practitioners usable guidance on when to partition in space versus feature space.","major_comments":[{"comment":"The central claim (abstract and §§III–IV) that spatial-domain JSCC “excels at handling random localized losses” while feature-domain JSCC “provides superior robustness to distributed erasures and maintains fidelity under low-loss scenarios” rests on mismatch curves (Figs. 2 and 5) whose operating points are not rate-matched. The manuscript itself notes that each model settles at a different BPP (Fig. 2: 0.3558–0.4005; Fig. 5: 0.2959–0.3558) under a single λ=2048. Because the entropy model can freely allocate bits to redundancy, the observed PSNR gaps may partly reflect unequal bit budgets rather than intrinsic spatial-continuity versus semantic-redundancy advantages. A rate-matched comparison (or an RD sweep at fixed erasure rates) is needed to isolate the claimed complementarity.","section":null},{"comment":"§III.B and Fig. 4 define the two-level importance map by protecting central face/upper-body blocks (ε=0.05) more than periphery (ε=0.1). This axiom is reasonable for talking-head VCD but is never validated against an alternative importance map or against a non-face-centric task. Without such a control, the claim that the non-uniform scheme is “semantic-guided” remains tied to a single, hand-crafted prior and does not generalize beyond the evaluated scenarios.","section":null}],"minor_comments":[{"comment":"Fig. 1 caption and surrounding text refer to a “multi-level block erasure interface” but never state the precise number of importance levels used at training time for the feature-domain progressive schedule; a short clarifying sentence would help.","section":null},{"comment":"The placeholder value -1 is justified by the [0,1] latent range, yet the manuscript does not discuss whether this choice interacts with the subsequent 8-bit quantization or entropy model; a brief note would remove ambiguity.","section":null},{"comment":"BPP values are reported inconsistently across figures (some captions list them, others do not). Adding a compact table of clean-channel BPP for every trained model would improve readability.","section":null},{"comment":"Typographical inconsistencies appear in author affiliations and in the arXiv header date; these are easily fixed.","section":null}],"recommendation":"major_revision","confidential_remarks":"The rate-matching gap is the single load-bearing experimental weakness; once addressed (even with a modest additional RD sweep), the paper becomes a solid contribution for a communications or multimedia journal. Scope is appropriate for eess.IV / JSAC-style venues."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The useful takeaway is a clean experimental comparison of two ways to partition a DVC-style neural video codec for multi-level block erasures: spatial blocks (16 local patches) versus feature-channel blocks (8 groups). They train with both uniform and two-level non-uniform erasures (central face protected more than periphery on talking-head video) and show the expected complementary behavior—spatial recovery looks like inpainting and handles localized losses well; feature recovery looks like semantic imputation and is more robust when losses are distributed or light.\n\nWhat is actually new is the systematic side-by-side on video, the semantic-guided two-level scheme, and the progressive top-k block decoding curves under non-uniform training. The mismatch plots (Figs. 2 and 5), the visual recovery sequences (Figs. 3 and 8), the regional PSNR split (Fig. 4), and the hierarchical decoding curve (Fig. 7) all line up with the story they tell. Training details are concrete (Vimeo-90k, λ=2048, VCD evaluation), and they are honest that the curves are not rate-matched.\n\nThat honesty is also the soft spot. Because each model settles at a different BPP under the same λ, the PSNR gaps that support “spatial excels at localized, feature at distributed/low-loss” are not fully isolated from rate and redundancy effects. The stress-test note is right on this; it is not fatal for an engineering paper, but it does mean the strongest abstract claim is a bit stronger than the isolation of the experiment. Secondary limits are the talking-head-only test set and the hand-crafted central/peripheral importance map—reasonable for the target scenario, not yet general. No code release.\n\nThis is for people building neural JSCC or unequal-protection video systems who want concrete trade-off numbers rather than theory. The math is standard, the data are consistent with the claims, and the citations are appropriate. I would send it to peer review; a referee can ask for rate-matched ablations and a broader dataset without the paper collapsing. Worth reading if you work in the area; I would cite the comparative figures if I were writing on erasure-aware video JSCC.","headline":"Solid empirical map of spatial vs feature partitioning for neural video JSCC under block erasures; the complementarity claim is useful but not cleanly rate-isolated.","tokens_in":9475,"tokens_out":544,"would_cite":true,"duration_ms":6245,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Spatial and feature block partitions give complementary defenses for neural video over erasures.","keywords":["block erasure channels","joint source-channel coding","semantic communication","video compression","neural JSCC","spatial-domain partitioning","feature-domain partitioning","unequal error protection"],"falsifier":"Retrain and re-test on a non-talking-head video set (sports, surveillance, or multi-object scenes) under the same two-level spatial erasure schedule; if the protected-region PSNR advantage disappears or reverses, the semantic-importance ranking fails.","tokens_in":9478,"feed_emoji":"📡","tokens_out":741,"duration_ms":8566,"temperature":0.7,"pith_summary":"Video traffic needs both compression and resilience when wireless links drop whole blocks of data. This paper builds a neural joint source-channel coder on a deep video compression backbone and splits the representation two ways: into spatial tiles of the motion and residual maps, or into channel groups of the latent features. Training with erasures teaches the decoder either to inpaint missing tiles from neighbors or to impute missing feature groups from semantic redundancy across channels. Experiments on talking-head conference video show the spatial design recovers random localized losses well, while the feature design stays more faithful under distributed losses and under low erasure rates. The work therefore supplies a practical map of when each partition style is preferable and how to protect face-centric regions more than background.","feed_headline":"Two block styles give video different erasure defenses","feed_subtitle":"Spatial tiles fix local losses; feature groups fix scattered ones and keep quality high at low loss.","key_machinery":"Semantic-aware neural JSCC that maps motion and residual latents into K blocks, each sent over an independent erasure channel whose probability can be uniform or importance-weighted, with erased blocks replaced by a fixed -1 placeholder so the decoder remains dimensionally fixed.","core_discovery":"When a neural video compressor is trained end-to-end against multi-level block erasures, spatial-domain partitioning recovers random localized losses by inpainting, whereas feature-domain partitioning recovers distributed losses by semantic imputation across channels, and the two styles trade spatial continuity against semantic redundancy.","pith_inferences":["Hybrid spatial-plus-feature partitions, or adaptive switching between them based on measured channel statistics, are a natural next architecture.","The same importance-weighted training could be driven by any differentiable task metric (object detection, action recognition) rather than face-centric PSNR.","Because the decoder learns to treat -1 as a reliable erasure flag, the scheme can be layered under existing packet-based transport without changing the physical layer.","The observed rate–robustness trade-off suggests an explicit multi-objective curriculum that jointly optimizes BPP and erasure resilience rather than fixing λ."],"forward_implications":["Designers can choose spatial partitioning when losses are expected to be localized and feature partitioning when losses are scattered.","Two-level non-uniform erasure can be used to give higher protection to face and upper-body regions without raising overall rate.","Feature-domain progressive training yields graceful degradation even when test erasure rates exceed the training schedule by an order of magnitude.","Under tight bandwidth only the top-ranked feature blocks need be sent, still yielding usable reconstruction quality.","The same training recipe can be adapted to other low-latency video applications that cannot retransmit."],"fun_headline_variants":["Spatial tiles inpaint local video erasures; features impute scattered ones","Two partitioning styles trade continuity for semantic recovery in video JSCC","Neural video JSCC: spatial blocks fix local losses, features fix distributed ones","Feature-domain partitioning recovers distributed erasures by semantic imputation","Spatial-domain JSCC excels at random local losses via inpainting"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The claim that protecting the central face blocks more than the periphery correctly captures what is semantically important, and that this ranking generalizes beyond the talking-head conference videos used for evaluation.","fun_headline_variants_meta":{"raw":{"variants":["Spatial tiles inpaint local video erasures; features impute scattered ones","Two partitioning styles trade continuity for semantic recovery in video JSCC","Neural video JSCC: spatial blocks fix local losses, features fix distributed ones","Feature-domain partitioning recovers distributed erasures by semantic imputation","Spatial-domain JSCC excels at random local losses via inpainting"]},"model":"grok-4.5","effort":"low","cost_usd":0.00416,"raw_usage":{"total_tokens":1180,"prompt_tokens":682,"num_sources_used":0,"completion_tokens":75,"cost_in_usd_ticks":41600000,"prompt_tokens_details":{"text_tokens":682,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":423,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":682,"tokens_out":75,"duration_ms":4309,"temperature":1.0,"reasoning_tokens":423,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-10T17:22:02.302415+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Retrain and re-test on a non-talking-head video set (sports, surveillance, or multi-object scenes) under the same two-level spatial erasure schedule; if the protected-region PSNR advantage disappears or reverses, the semantic-importance ranking fails.","supporting_citations":[],"review_version":1}