{"id":"56aae4da-de07-4d51-b78d-6e54888b9b30","arxiv_id":"2506.07391","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"D-NTSC and D-NTSCC, built on nonlinear transform coding with joint entropy modeling and spatial alignment, outperform existing distributed image transmission baselines on KITTI and Cityscapes.","lead":"This paper proposes two deep-learning schemes, D-NTSC and D-NTSCC, for transmitting correlated camera images to a central receiver by explicitly modeling their joint statistics with nonlinear transform coding and aligning latent features with learned geometric transforms. Experiments on stereo datasets show both schemes outperform several existing distributed image transmission baselines in rate-distortion and perceptual quality.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported hyperprior rate H(z1,z2) is not achievable by the separate arithmetic encoders in Figs. 2-3; without Slepian-Wolf binning, the actual rate is near H(z1)+H(z2), which would shift the RD curves and may erase the claimed gains.","rationale":"The reader's weakest assumption concerned the conditional independence y1 - z1 - z2 - y2. That is a modeling choice and is not the primary threat: even if that Markov condition held exactly, the joint entropy H(z1,z2) of the hyperpriors cannot be achieved by two separate arithmetic encoders without binning. The paper defines the hyperprior transmission cost as the joint entropy and splits it equally between users, but the architecture provides no mechanism (Slepian-Wolf syndromes, distributed arithmetic coding, or a conditional code using decoded side information) to realize this rate. This is a correctness issue in the reported rate accounting, not merely a modeling approximation. It directly affects the central empirical claim: if the actual hyperprior rate is H(z1)+H(z2), the RD curves shift right, and the claimed gains over LDMIC, DHF-JSCC, and the independent-entropy ablation may shrink or disappear. The concern is concrete and testable by measuring actual bitstream lengths. I recommend CONDITIONAL because the paper could be repaired by implementing an actual distributed source coding scheme for the hyperpriors or by honestly reporting the achievable separate-encoding rate; if the measurements show the rate gap is large, the verdict should move toward reject.","tokens_in":18484,"tokens_out":28009,"duration_ms":343679,"concrete_test":"Run the trained D-NTSC model on the KITTI Stereo test set and compute the actual bitstream lengths produced by the two independent arithmetic encoders using the marginal models p(¯z1) and p(¯z2). Recompute the rate-distortion curves in Fig. 9 and the joint-vs-independent entropy model ablation in Fig. 10 using these actual hyperprior bits instead of 1/2 H(¯z1,¯z2). If the actual per-user rate exceeds the reported R_i by more than a few percent, or if the 8.43% bitrate saving from the joint entropy model disappears when measured with achievable distributed coding, the central claim of rate-distortion advantage is invalid.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The rate-distortion results for both D-NTSC and D-NTSCC depend critically on the hyperprior cost being the joint entropy H(¯z1,¯z2), as stated in Eq. (7), with each user charged 1/2 H(¯z1,¯z2) in Eqs. (8) and (9). However, the system architecture uses two physically separated transmitters, each with its own arithmetic encoder (the 'AE' blocks in Fig. 2 and Fig. 3). Separate encoders cannot achieve the joint entropy of two correlated sources without an explicit distributed source coding mechanism such as Slepian-Wolf binning or distributed arithmetic coding. The paper does not describe or implement any such mechanism; Section IV-C only mentions that each transmitter can use the joint model to form an MMSE estimate z2* of the other hyperprior for rate allocation, which is not a method for jointly entropy-coding z1 and z2. Consequently, the actual hyperprior bitstream length produced by the two independent arithmetic coders would be approximately H(z1)+H(z2), not H(z1,z2). This means the reported bit rates in Figs. 9-13 are optimistic lower bounds. The claimed bitrate savings, including the 30.7% reduction over LDMIC and the 8.43% gain of the joint entropy model in the ablation, may be partially or wholly an artifact of this unachievable rate accounting rather than a real coding gain.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes D-NTSC and D-NTSCC, two learning-based distributed image transmission schemes for correlated sources, corresponding to separate source-channel coding (SSCC) and joint source-channel coding (JSCC). Both schemes use nonlinear transform coding with Swin Transformer backbones, a joint entropy model for the hyperpriors z1 and z2, and a spatial transformation module that warps one latent representation to align with the other before joint synthesis decoding. Loss functions are derived from variational inference, with rate terms including the claimed joint entropy H(z̄1,z̄2) and per-user distortion terms. Experiments on KITTI Stereo and Cityscapes report state-of-the-art rate-distortion and channel-rate performance relative to learning-based distributed coding baselines such as LDMIC, D-DJSCC, and DHF-JSCC, plus ablations for the joint entropy model and the transformation module.","tokens_in":18785,"tokens_out":6520,"duration_ms":74024,"significance":"If the reported rate-distortion gains are real, the paper makes a useful contribution by moving distributed image semantic communication from purely data-driven correlation learning toward an explicit probabilistic model of source correlation, and by introducing a geometric alignment module that is broadly applicable. The experimental section is broad, includes two datasets and multiple metrics, and the ablation studies isolate the joint entropy model and the transformation module. The promise to release code is a welcome step for reproducibility. However, the core theoretical and practical claims rest on two load-bearing assumptions that are not adequately supported: the achievability of the joint hyperprior entropy by two separate arithmetic encoders, and a Markov-chain conditional-independence structure for the latent representations. These issues directly affect the numerical bit-rate and bandwidth-savings claims, so the significance of the empirical gains cannot be assessed until they are resolved.","major_comments":[{"comment":"The reported hyperprior rate of (1/2)H(z̄1,z̄2) per user is not achievable by the architecture described. Each transmitter runs its own arithmetic encoder on its own z̄_i, and the paper does not specify or implement any distributed source coding mechanism, such as Slepian-Wolf binning, syndrome-based coding, or distributed arithmetic coding. The joint model in Eq. (16) is used only for rate estimation and for providing the MMSE estimate z2* of the other user's hyperprior for rate allocation (Section IV-C, Remarks), which does not shorten the entropy-coded bitstream. The actual rate produced by two independent arithmetic encoders is approximately H(z̄1)+H(z̄2), not H(z̄1,z̄2). Consequently, Eqs. (8) and (9) understate the true bit rate by H(z̄_i)-1/2H(z̄1,z̄2) per user. This affects all reported rate-distortion curves (Figs. 9-13) and specifically the claimed 8.43% bit-rate saving of the joint entropy model in Fig. 10 and the 30.7% bit-rate reduction over LDMIC in Section VI-B. Please either implement a genuine distributed entropy-coding scheme for the hyperpriors or report all rates using the marginal entropy models that match the actual arithmetic coders.","section":"Section IV-C, Eqs. (7)-(9), and Figs. 2-3"},{"comment":"The factorization p(y1,y2|z1,z2)=p(y1|z1)p(y2|z2) is asserted to follow from the Markov chain y1 - z1 - z2 - y2, but this chain is neither proven nor empirically validated. The analysis transforms ga1 and ga2 are learned independently and the hyperpriors zi are low-dimensional summaries, so there is no evident reason that all correlation between y1 and y2 is mediated by z1 and z2. If residual correlation remains, Eq. (10) is not the correct variational objective and the model may underestimate the joint rate. I recommend an empirical test on held-out data: compare the model's rate -log p(y1|z1)-log p(y2|z2)-log p(z1,z2) with the actual bitstream length produced by the separate arithmetic encoders, or fit a richer conditional model and report the rate difference. The analogous Markov chain ŝ1 - z1 - z2 - ŝ2 in Appendix B raises the same concern for D-NTSCC.","section":"Appendix A, Eqs. (20)-(22), and Proposition 1"},{"comment":"The bandwidth allocation in Eq. (18) is a heuristic nearest-neighbor discretization of -η log p(y^j_1|z̄1) onto the finite set V, and the paper does not analyze or ablate this approximation. Since the rate estimates used in this allocation are themselves affected by the unachievable hyperprior cost in Eq. (9), it is unclear how much of the reported bandwidth savings in Figs. 11-12 comes from the coding scheme and how much from the optimistic rate accounting. Please provide results with the corrected rate definition and, if possible, an ablation of the discretization granularity.","section":"Section VI-C and Eq. (18)"}],"minor_comments":[{"comment":"The section heading contains a typo: 'BACKGROND' should be 'BACKGROUND'.","section":"Section II heading"},{"comment":"The notation N(m_k, Σ_k) * U([-1/2,1/2]^2) should specify that this is a two-dimensional convolution of the Gaussian density with the bivariate uniform density; the current notation is easy to misread as a product of densities.","section":"Eq. (16)"},{"comment":"The AE/AD blocks are drawn separately for each user's hyperprior, which visually reinforces the concern that the two hyperpriors are entropy-coded independently; if joint decoding of z1 and z2 is intended, the figures should show the side-information path.","section":"Fig. 2 and Fig. 3"},{"comment":"The complexity comparison would benefit from stating the input image resolution and channel-use count used for the FLOPs measurements, since these quantities strongly affect the comparison.","section":"Section VI-F, Table I"},{"comment":"The phrase 'both transmitters share this joint probability function, they can predict each other's transmission rate' is about rate estimation, not about joint entropy coding; the distinction should be stated explicitly to avoid implying that the joint model is used by the arithmetic encoders.","section":"Section IV-C, Remarks"}],"recommendation":"major_revision","confidential_remarks":"The central empirical claims are interesting, but the hyperprior rate accounting must be fixed before the results can be interpreted. The Markov-chain assumption in the variational derivation is also unproven and should be validated empirically. These are fixable within the manuscript's scope, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know about this paper if you follow learned image compression or distributed JSCC. It is a serious attempt to bring nonlinear transform coding to multi-view distributed transmission, and the experimental work is real. But there is a load-bearing flaw in how the bit rate is counted.\n\nThe core ideas are a joint bivariate entropy model for the hyperpriors and a learned homography-based transformation module that aligns one view's latents to the other before joint synthesis. Both are tested with ablations and show consistent gains on KITTI Stereo and Cityscapes. The Swin Transformer backbone is a reasonable choice, and the complexity comparison is honest.\n\nThe main problem: equations (7) and (8) charge each user 1/2 H(z1,z2) for the hyperpriors, assuming the two separate arithmetic encoders in Figures 2 and 3 achieve the joint entropy of correlated sources. They do not, absent a Slepian-Wolf binning or distributed arithmetic coding scheme, and the paper describes none. In practice the bitstream cost will be near H(z1)+H(z2), so the reported bit rates are optimistic lower bounds. The 30.7% bit-rate reduction vs. LDMIC and the 8.43% joint-model ablation gain may be partly an artifact of that accounting. This is not a minor point; the rate axis is the independent variable in all the rate-distortion curves.\n\nThe variational derivation also uses an unproven Markov chain y1 - z1 - z2 - y2 to drop a term. It may be a reasonable modeling assumption, but it needs at least an empirical check, and the paper does not provide one. Minor issues: no error bars, and code is promised but not yet available.\n\nWho it is for: people working on distributed image compression or learned JSCC will find the architecture and ablation methodology useful, but the headline rate-distortion numbers should be treated with suspicion until the rate accounting is corrected or a distributed entropy coder is implemented.\n\nRecommendation: send it to peer review, but reviewers should push hard on the rate accounting. If the authors fix that or add a mechanism to achieve the joint entropy, the paper could be solid. As it stands, the central claim is overstated.","headline":"Strong engineering contribution to distributed image coding undermined by an unachievable rate accounting for the hyperpriors.","tokens_in":19361,"tokens_out":2137,"would_cite":false,"duration_ms":20978,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that explicitly modeling the correlation between two camera views—probabilistically through a joint entropy model and geometrically through a spatial transformer—lets distributed image transmission beat purely data-driven…","keywords":["distributed source coding","joint source-channel coding","nonlinear transform coding","semantic communication","joint entropy model","spatial transformer","multi-view image compression","variational inference"],"falsifier":"Train the D-NTSC encoder on a stereo dataset, then compute the empirical conditional mutual information $I(\\tilde y_1;\\tilde y_2 \\mid \\tilde z_1, \\tilde z_2)$ on a held-out set. If the estimate is clearly nonzero, the joint entropy model fails to capture all source correlation and the derived loss is not the true variational objective.","tokens_in":18238,"feed_emoji":"📡","tokens_out":4908,"duration_ms":45258,"temperature":0.7,"pith_summary":"The paper claims that distributed image transmission can be improved by modeling the correlation between two cameras' views explicitly, instead of letting a neural network learn it implicitly. To that end it builds two coding schemes on nonlinear transform coding (NTC): D-NTSC for separate source and channel coding and D-NTSCC for joint source-channel coding. Both use a learned joint entropy model of the two latent representations to allocate rate adaptively and a spatial transformer to align the two views before joint decoding. On KITTI Stereo and Cityscapes, the schemes are reported to achieve state-of-the-art rate-distortion results, with D-NTSC reducing bit rate by roughly 30% versus a leading distributed coding baseline and D-NTSCC saving up to 50% bandwidth over a distributed deep JSCC baseline. The broader point is that coupling variational autoencoders with explicit geometric alignment can turn distributed source coding into a practical, learnable system.","feed_headline":"Modeling source correlation explicitly boosts distributed image coding","feed_subtitle":"Two NTC-based codecs beat distributed SSCC and JSCC baselines on stereo wireless transmission.","key_machinery":"The argument turns on three components. (1) A joint entropy model: the hyperpriors $z_1,z_2$ are modeled jointly by a pairwise factorized bivariate Gaussian mixture, giving a density $p(\\tilde z_1,\\tilde z_2)$ that approximates the true joint distribution of the latents and guides adaptive rate allocation for both users. (2) A transformation module based on spatial transformer networks, which estimates a projective (homography) matrix $M_i \\in \\mathbb{R}^{3\\times 3}$ and warps the other user's latent representation into the expected viewpoint before the joint synthesis transform, maximizing usable side information. (3) A variational-inference loss derivation in which the Markov chain $\\tilde y_1 - \\tilde z_1 - \\tilde z_2 - \\tilde y_2$ lets the KL objective decompose into distortion plus rate terms, with the joint hyperprior density replacing an independence assumption. The Swin Transformer backbone supplies the nonlinear analysis and synthesis transforms in both schemes.","core_discovery":"For two correlated image sources separately encoded at distinct transmitters and jointly decoded at a common receiver, the paper establishes that explicitly modeling their joint distribution in latent space yields measurable gains over purely data-driven correlation learning. The central result is the joint entropy model $p(\\tilde z_1,\\tilde z_2)$ built as a pairwise factorized bivariate Gaussian mixture, which lets each transmitter estimate the other's rate and drives adaptive rate allocation, together with a projective transformation module that warps one latent representation toward the other's viewpoint before decoding. Variational inference produces the loss functions (12) and (13), whose rate terms factor through the Markov chain $\\tilde y_1 - \\tilde z_1 - \\tilde z_2 - \\tilde y_2$, with the joint hyperprior density replacing the independence assumption used by prior distributed JSCC work. Experiments report D-NTSC outperforming LDMIC by 0.3--0.6 dB PSNR at comparable bit rates and reducing bit rate by 30.7% on KITTI Stereo and 26.0% on Cityscapes, while D-NTSCC saves up to 50% bandwidth over D-DJSCC and about 30% over DHF-JSCC at equal quality. An ablation attributes part of the gain to the joint entropy model itself, which reduces bit rate by up to 8.43% relative to an independent entropy model.","pith_inferences":["The Markov-chain assumption could be tested directly: estimate the empirical conditional mutual information $I(\\tilde y_1;\\tilde y_2 \\mid \\tilde z_1, \\tilde z_2)$ on the learned latents; if it is clearly nonzero, the joint entropy model undercounts the true rate and the derived loss is only an approximation to the variational objective.","The framework likely transfers to other multi-view tasks such as multi-camera surveillance or light-field compression, where pose alignment and joint entropy are natural priors, but the paper evaluates only standard stereo image pairs, so the transfer remains untested.","The bandwidth-savings numbers compare against baselines with different architectures; a controlled comparison using the same Swin Transformer backbone with and without the joint entropy model would isolate the contribution of explicit probabilistic correlation modeling.","The rate-estimation procedure assumes the channel conditions used in training; performance under SNR mismatch or fading channels is not examined, and that is where the JSCC scheme's adaptive rate mechanism would face a realistic stress test."],"forward_implications":["D-NTSC and D-NTSCC give system designers two operating points: choose D-NTSCC for low rates and low signal-to-noise ratios, and D-NTSC for higher rates where separate source and channel coding performs better.","The joint entropy modeling can be extended to more than two users by generalizing the bivariate Gaussian mixture to multivariate Gaussian variables, as the paper notes.","Because each transmitter can estimate the other's coding rate from the shared joint probability model, adaptive rate allocation and resource coordination can happen without a central encoder.","The probabilistic correlation model and the geometric alignment module are separable components that could be combined with other multi-view coding or transmission pipelines.","The reported gains suggest that explicit correlation structure—both probabilistic and geometric—is a stronger inductive bias than implicit attention-based correlation learning for distributed image coding."],"supporting_citations":[{"why":"Supplies the scale-hyperprior variational autoencoder formulation and the conditional Gaussian entropy model that D-NTSC extends to two correlated sources.","marker":"[29]"},{"why":"Supplies the point-to-point nonlinear transform source-channel coding (NTSCC) framework that D-NTSCC generalizes to the distributed setting.","marker":"[7]"},{"why":"Establishes the information-theoretic Slepian-Wolf principle that separate encoding with joint decoding can achieve joint-encoding rates, motivating the system model.","marker":"[16]"},{"why":"LDMIC is the state-of-the-art distributed multi-view image coding baseline that D-NTSC must beat, providing the main source-coding comparison.","marker":"[22]"},{"why":"D-DJSCC is the distributed deep JSCC baseline that D-NTSCC compares against, representing implicit cross-attention-based correlation learning.","marker":"[23]"},{"why":"DHF-JSCC is the recent distributed JSCC baseline whose independent hyperprior entropy model is replaced by the proposed joint model, making it the direct architectural competitor.","marker":"[24]"},{"why":"Supplies the Swin Transformer backbone used for the nonlinear analysis and synthesis transforms in both D-NTSC and D-NTSCC.","marker":"[30]"},{"why":"Supplies the spatial transformer network mechanism that underlies the proposed transformation module for aligning latent representations.","marker":"[31]"}],"fun_headline_variants":["Explicit latent correlation model improves distributed image coding","Joint entropy model beats data-driven correlation in distributed JSCC","NTC-based codecs cut bandwidth by half for stereo wireless transmission","Distributed codecs leverage variational inference for joint source modeling","Pairwise joint entropy model drives adaptive rate allocation in coding"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The two images' latent representations are assumed to be independent once their hyperpriors are known, so any correlation left between the latents after conditioning on the hyperpriors is invisible to the rate model and would make the reported bit rates underestimates.","fun_headline_variants_meta":{"raw":{"variants":["Explicit latent correlation model improves distributed image coding","Joint entropy model beats data-driven correlation in distributed JSCC","NTC-based codecs cut bandwidth by half for stereo wireless transmission","Distributed codecs leverage variational inference for joint source modeling","Pairwise joint entropy model drives adaptive rate allocation in coding"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000762,"raw_usage":{"total_tokens":3432,"prompt_tokens":1045,"completion_tokens":2387,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":661,"completion_tokens_details":{"reasoning_tokens":2306}},"tokens_in":661,"tokens_out":2387,"duration_ms":18611,"temperature":1.0,"reasoning_tokens":2306,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T05:35:23.672752+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the D-NTSC encoder on a stereo dataset, then compute the empirical conditional mutual information $I(\\tilde y_1;\\tilde y_2 \\mid \\tilde z_1, \\tilde z_2)$ on a held-out set. If the estimate is clearly nonzero, the joint entropy model fails to capture all source correlation and the derived loss is not the true variational objective.","supporting_citations":[{"cited_title":"Noiseless coding of correlated information sources,","cited_arxiv_id":null,"evidence_quote":"Establishes the information-theoretic Slepian-Wolf principle that separate encoding with joint decoding can achieve joint-encoding rates, motivating the system model."},{"cited_title":"LDMIC: Learning-based Distributed Multi-view Image Coding","cited_arxiv_id":"2301.09799","evidence_quote":"LDMIC is the state-of-the-art distributed multi-view image coding baseline that D-NTSC must beat, providing the main source-coding comparison."},{"cited_title":"Distributed image transmission using deep joint source-channel coding,","cited_arxiv_id":null,"evidence_quote":"D-DJSCC is the distributed deep JSCC baseline that D-NTSCC compares against, representing implicit cross-attention-based correlation learning."},{"cited_title":"Content adaptive distributed joint source-channel coding for image transmission with hyperprior,","cited_arxiv_id":null,"evidence_quote":"DHF-JSCC is the recent distributed JSCC baseline whose independent hyperprior entropy model is replaced by the proposed joint model, making it the direct architectural competitor."},{"cited_title":"Swin transformer: Hierarchical vision transformer using shifted windows,","cited_arxiv_id":null,"evidence_quote":"Supplies the Swin Transformer backbone used for the nonlinear analysis and synthesis transforms in both D-NTSC and D-NTSCC."},{"cited_title":"Spatial transformer networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the spatial transformer network mechanism that underlies the proposed transformation module for aligning latent representations."}],"review_version":1}