{"id":"fa569626-d5bf-4966-8079-98334a35224a","arxiv_id":"2411.10346","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"BiDense uses an input-adaptive binarizer and a channel-adaptive full-precision bypass to make 1-bit neural networks accurate on semantic segmentation and monocular depth estimation.","lead":"This paper presents BiDense, a binary neural network for pixel-level tasks like segmentation and depth estimation, using two tricks to keep more full-precision information: an input-adaptive binarizer and full-precision channel bypasses. It reports near-full-precision accuracy on four benchmarks at a small fraction of the memory and compute, though the strongest comparison uses a randomly initialized full-precision model.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"BiDense's headline 'comparable to full-precision' is tested only against a randomly initialized FP32 baseline; against the paper's own pretrained ConvNeXt*, the gaps are large (18.75 vs 40.96 mIoU on ADE20K), so the central accuracy claim is unsupported as stated.","rationale":"The paper has a real methodological contribution: DAB and CFB improve over prior BNN baselines on both tasks, and the ablation in Table 4 supports the importance of both components. That part of the argument is internally consistent. The problem is the headline claim. The abstract states that BiDense 'achieves performance levels comparable to full-precision models,' but the experiments compare BiDense to the randomly initialized ConvNeXt, not the ImageNet-pretrained ConvNeXt* that is the usual reference for dense prediction. The paper itself reports ConvNeXt* with 40.96 mIoU on ADE20K versus BiDense's 18.75; that is a 54% relative shortfall, not 'comparable.' The same pattern appears in depth: on NYUv2, AbsRel 0.189 vs 0.117; on KITTI, 0.095 vs 0.065. The random-init baseline (20.39 mIoU; 0.201 and 0.092 AbsRel) is the only full-precision model BiDense approaches. Unless the paper restricts its claim to 'randomly initialized full-precision models trained on the target set,' the central accuracy assertion is unsupported. A separate red flag is Table 2: ConvNeXt* on PASCAL VOC is reported with mIoU 99.72 and pixAcc 98.58, an implausible result for UPerNet/ConvNeXt-Tiny; the VOC comparison should be rerun. Because this concern is about claim calibration rather than the validity of the method, the appropriate disposition remains CONDITIONAL (major revision): require the authors to compare against or at least clearly report the pretrained FP32 baseline, and to correct the VOC numbers. The reader's weakest assumption was about CFB's efficiency overhead; though that is also a real gap (no latency/memory measurements), the accuracy-baseline issue is more directly contradicted by the paper's own tables, so I treat it as the load-bearing concern.","tokens_in":14997,"tokens_out":14787,"duration_ms":140494,"concrete_test":"Recompute the headline comparisons using the ImageNet-pretrained ConvNeXt* row (Tables 1-3) as the full-precision reference: if the ADE20K mIoU gap remains above 20 points or the NYUv2 AbsRel gap remains above 0.05, then the phrase 'comparable to full-precision models' is unsupported and should be replaced with 'models trained from random initialization.' Separately, rerun the official VOC evaluation for the ConvNeXt* row; if mIoU is not 99.72, correct Table 2 and re-evaluate the VOC comparisons.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Tables 1-3 evaluate the central 'comparable to full-precision models' claim against ConvNeXt, the randomly initialized FP32 baseline, not ConvNeXt*, the ImageNet-pretrained baseline. This matters because dense prediction models are normally initialized from ImageNet, and the pretrained baseline is much stronger: ADE20K mIoU 40.96 vs 20.39; NYUv2 AbsRel 0.117 vs 0.201; KITTI AbsRel 0.065 vs 0.092. BiDense is close to the random baseline (18.75 mIoU, AbsRel 0.189 and 0.095) but far from the pretrained one. The abstract promises comparability to full-precision models without this qualifier. In addition, the pretrained VOC row (ConvNeXt*: pixAcc 98.58, mIoU 99.72) is implausible for UPerNet/ConvNeXt-Tiny on PASCAL VOC and should be rechecked before any comparison on that dataset is used. The accuracy half of the central claim therefore rests on an atypical reference point; the paper must either compare against the standard pretrained baseline or narrow the claim to 'models trained from random initialization.'","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes BiDense, a binary neural network (BNN) framework for dense prediction tasks, specifically semantic segmentation and monocular depth estimation. BiDense introduces two components: a Distribution-adaptive Binarizer (DAB), which computes binarization thresholds and scaling factors from the mean and mean absolute deviation of the input activations, and a Channel-adaptive Full-precision Bypass (CFB), which provides full-precision skip paths when a binary convolution changes the number of channels. The authors evaluate BiDense against several prior BNN methods on ADE20K, PASCAL VOC, NYUv2, and KITTI, and report that it outperforms existing BNNs and approaches the accuracy of full-precision models at a fraction of the parameters and operations. The paper also includes ablations, an extension to a DPT-style hybrid architecture, and visualizations of information entropy.","tokens_in":15231,"tokens_out":6418,"duration_ms":61640,"significance":"If the central claim were fully supported, BiDense would be the first binary network to closely match full-precision accuracy on both semantic segmentation and monocular depth estimation, which would be a meaningful advance for efficient dense prediction. The two proposed components are simple and potentially generalizable, and the ablation study (Table 4) shows that the CFB contributes a large accuracy gain (mIoU from 9.78 to 18.75 on ADE20K). The paper also provides a public code link and reports comparisons against multiple BNN baselines in a unified experimental setup. However, the headline claim of 'comparable to full-precision' is only established against a randomly initialized full-precision baseline, not the standard ImageNet-pretrained baseline, and the OPs definition in Section 4 is ambiguous. These issues materially weaken the significance as currently stated; the contribution is better characterized as a strong BNN for dense prediction trained from random initialization, with the full-precision comparison requiring substantial qualification.","major_comments":[{"comment":"The central claim that BiDense achieves 'performance levels comparable to full-precision models' is supported only by comparison with the randomly initialized FP32 baseline (ConvNeXt), not the ImageNet-pretrained baseline (ConvNeXt*) that the paper itself lists. On ADE20K, the gap to ConvNeXt* is 22.21 mIoU (18.75 vs 40.96), whereas the gap to the random ConvNeXt is only 1.64 mIoU (18.75 vs 20.39). Similarly, on NYUv2 AbsRel BiDense is 0.189 vs 0.117 for ConvNeXt* and 0.201 for random ConvNeXt, and on KITTI AbsRel it is 0.095 vs 0.065 for ConvNeXt*. Since dense prediction models in practice are initialized from ImageNet, the 'comparable to full-precision' statement in the abstract and the text in §4.1 and §4.2 is unsupported. The authors should either compare against the pretrained baseline for all methods and report those numbers prominently, or explicitly narrow the claim to 'models trained from random initialization.'","section":"Abstract, §4.1, Tables 1–3"},{"comment":"The ConvNeXt* row on PASCAL VOC reports pixAcc 98.58 and mIoU 99.72. An mIoU of 99.72 for ConvNeXt-Tiny with UPerNet on PASCAL VOC is implausible, and the mIoU being higher than pixAcc is anomalous for a 21-class segmentation benchmark. This appears to be a data error or a typo, but it is used in the running text to assert that BiDense 'approaches the FP32 baseline.' The row must be corrected or removed before any claims about VOC are made; as printed, it undermines the credibility of the comparison.","section":"Table 2"},{"comment":"The computational-cost formula at the start of §4 is internally inconsistent with the reported numbers. If OPs_f denotes the full-precision model's total operations, then OPs = OPs_b + OPs_f with OPs_b = OPs_f/64 would give roughly 1.016×OPs_f for any BNN, contradicting Table 1, where BiDense is listed at 5.37 G vs 288.37 G for the FP32 model. The formula only makes sense if OPs_f is reinterpreted as the remaining full-precision operations inside the binary network, but that interpretation is never stated. Furthermore, the 'significantly reducing computational costs' claim rests entirely on this FLOPs proxy: no latency, energy, or memory measurements are provided, and the CFB bypass carries full-precision activations through the network, whose memory footprint is not accounted for. Please clarify the definition and, ideally, add at least one hardware measurement (latency or energy) to support the efficiency claim.","section":"§4, OPs definition"},{"comment":"The fusion-up procedure is not fully specified. Eq. (13) defines K = floor(Cin / (Cout mod Cin)), which divides by zero when Cout is an exact multiple of Cin — a common case in dense prediction architectures. The text also does not specify how the 'remainder' channels are merged when the remainder is zero, nor how the pooled output is concatenated with the repeated channels in that case. Please provide the complete algorithm, including the zero-remainder case, so that the implementation is reproducible without recourse to the code repository.","section":"§3.3, Eqs. (12)–(13)"}],"minor_comments":[{"comment":"The approximation 'lim_{x→1} exp(a(x−1)) ≈ ax' is inaccurate; the correct first-order approximation near x=1 is exp(a(x−1)) ≈ 1 + a(x−1), not ax. This does not affect the experiments but should be corrected.","section":"§3.2, Eq. (10)"},{"comment":"The sentence 'only 3.56% and 1.64% reductions of pixAcc and mIoU' compares BiDense to ConvNeXt, the random baseline, without saying so. Please state explicitly which baseline is being used in each comparison, since the pretrained ConvNeXt* has much larger gaps.","section":"§4.1"},{"comment":"The supplementary's limitation statement that DAB is not suited for transformer/self-attention modules is important and should be in the main text, because it materially qualifies the claim of a 'generalized' binary framework for dense prediction.","section":"Supplementary, §A and §C"},{"comment":"Reference [3] is cited as BiSRNet in the text, but the listed title is 'Binarized spectral compressive imaging,' which does not match a semantic segmentation method called BiSRNet. Please verify that the correct reference is cited.","section":"References"},{"comment":"Setting α(xs) = ||xs||_1 / n_xs yields a dramatic drop to 4.97 mIoU, which suggests training instability. It would be helpful to report the training configuration for this variant (same learning rate, epochs, seeds) to confirm the comparison is not affected by a single unlucky run.","section":"Table 4, item 7"}],"recommendation":"major_revision","confidential_remarks":"The implausible ConvNeXt* row in Table 2 and the ambiguous, seemingly inconsistent OPs formula in Section 4 give me pause about the rigor of the reported numbers. I would ask the editor to have the authors double-check all reported values, especially the pretrained-baseline rows, before considering the paper further. The accuracy claim against random initialization only is a significant reduction in scope from the abstract, and the paper should be revised to present the pretrained-baseline comparison as the main reference point."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's my read on BiDense.\n\nThe useful core is real: DAB is a sensible input-adaptive binarizer (mean-based threshold with learnable shift and a scaled MAD), and CFB generalizes BiSRNet's channel-mismatch bypass to arbitrary channel transformations with cheap pooling/repetition. The ablation shows both components matter, and the gains over prior BNN baselines are consistent across four datasets—ADE20K mIoU goes from 9.74 (BiSRNet) to 18.75, and PASCAL VOC from 47.5 to 75.2. That is a legitimate step for 1-bit dense prediction, and the first-depth-estimation claim looks plausible.\n\nThe problem is the packaging. The abstract says 'comparable to full-precision models,' but the only FP32 comparison close to BiDense is ConvNeXt with random initialization. Against the ImageNet-pretrained ConvNeXt*, the gaps are enormous (ADE20K: 18.75 vs 40.96 mIoU; NYUv2 AbsRel 0.189 vs 0.117). The paper does show both baselines in Tables 1 and 3, but then compares only to the random one when making the headline claim. That needs to be fixed—either compare to the standard pretrained baseline or narrow the claim.\n\nThere are two concrete technical flags. First, Table 2's ConvNeXt* row (pixAcc 98.58, mIoU 99.72 on PASCAL VOC) is implausible for UPerNet/ConvNeXt-Tiny; that number should be rechecked before anyone trusts the VOC comparison. Second, the CFB fusion-up formula (Eq. 13) divides by zero when Cout mod Cin = 0; there needs to be a conditional branch or a clearly stated convention. Both are fixable, but they undercut confidence in the implementation.\n\nThe OPs and memory accounting also deserve scrutiny. The paper counts total cost as OPs_f/64 + OPs_f, which includes the full-precision bypass at every conv. No latency or energy numbers are given, and the bypass carries full-precision activations through the network. The 1.86%-of-FP32 OPs number is computed on paper, not on hardware. For a paper that promises efficiency, that is a significant missing piece.\n\nBottom line: the method and experiments are worth referee time, but the framing overstates the result. This is a solid BNN-dense-prediction contribution in need of a revision that fixes the baseline comparison, the VOC anomaly, and the CFB edge case, and that reports real deployment costs.","headline":"Useful BNN module for dense prediction, but the 'comparable to full-precision' claim rests on a random-init baseline and one implausible table entry.","tokens_in":15817,"tokens_out":4929,"would_cite":true,"duration_ms":44239,"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":"BiDense claims that a binary neural network with adaptive binarization and full-precision bypasses can match full-precision accuracy on semantic segmentation and monocular depth estimation while using a small fraction of the memory and…","keywords":["binary neural networks","dense prediction","semantic segmentation","monocular depth estimation","model binarization","information entropy","full-precision bypass","adaptive binarization"],"falsifier":"Run BiDense on a mobile or edge GPU and measure end-to-end latency, peak memory, and energy against the FP32 baseline and a BiDense variant with the CFB removed. If the full-precision bypasses make runtime or memory approach the FP32 model, the paper's core efficiency claim fails.","tokens_in":14746,"feed_emoji":"⚡","tokens_out":9342,"duration_ms":78979,"temperature":0.7,"pith_summary":"BiDense aims to show that binary neural networks, which store weights and activations as single signs, can be made accurate enough for dense prediction — semantic segmentation and monocular depth estimation — where every pixel matters and earlier binarization methods lost too much detail. Its strategy is to keep real-valued information flowing through the network in two places: the binarizer itself, which adaptively sets thresholds and scales from each input's mean and mean absolute deviation, and a full-precision bypass that routes real-valued features around binary convolutions when the channel count changes. Across ADE20K, PASCAL VOC, NYUv2, and KITTI, BiDense outperforms previous binary networks and lands close to its full-precision ConvNeXt-UPerNet baseline, and on NYUv2 it reports even better depth metrics than that baseline. The paper also positions itself as the first binarization method applied to monocular depth estimation. If the results hold, pixel-level prediction on memory- and power-limited devices becomes a realistic target for 1-bit networks.","feed_headline":"1-bit network nears full-precision accuracy on segmentation and depth","feed_subtitle":"Adaptive thresholds and full-precision skips keep binary models within a few points of FP32 at about 2% of the compute.","key_machinery":"The load-bearing machinery is a redesigned binary convolutional layer with two additions. The Distribution-adaptive Binarizer (DAB) sets the binarization threshold as $\\beta(x)=k\\bar{x}+b$, with learnable $k$ and $b$ and $\\bar{x}$ the input mean, then assigns a per-input scale $\\alpha(x_s)=\\exp(a(\\|x_s\\|_{\\ell^1}/n_{x_s}-1))$, where $x_s=x-\\beta(x)$, blending a learnable global baseline with the current input's mean and mean absolute deviation. The Channel-adaptive Full-precision Bypass (CFB) implements fusion-down, 1D adaptive average pooling along the channel dimension with kernel and stride equal to avoid overlap, and fusion-up, channel repetition plus pooling of the remainder, so a full-precision copy of any layer input can be reshaped to the binary output's channel count and added to it. Together they let the binary convolution learn residuals on top of a real-valued signal, while the entropy argument from earlier binarization work explains why the adaptive threshold preserves more information.","core_discovery":"The paper's central claim is that the accuracy gap in binary dense prediction comes mostly from information discarded at two well-defined points, and that both can be patched cheaply. First, fixed binarization thresholds and scales, once trained, misfire on inputs whose distributions shift; the Distribution-adaptive Binarizer (DAB) replaces them with input-dependent values, $\\beta(x)=k\\bar{x}+b$ and $\\alpha(x_s)=\\exp(a(\\|x_s\\|_{\\ell^1}/n_{x_s}-1))$, so that binarization keeps the entropy of the 1-bit signal near its maximum. Second, encoder-decoder architectures change channel counts constantly, which blocks full-precision skip connections; the Channel-adaptive Full-precision Bypass (CFB)'s fusion-down and fusion-up resize a full-precision copy of the input to match the binary convolution's output shape, letting the binary path learn residuals while the real-valued path carries fine detail. With both mechanisms, the paper reports 67.25% pixel accuracy and 18.75% mIoU on ADE20K, close to the 70.81% and 20.39% of its randomly initialized FP32 baseline, and 75.2% mIoU on PASCAL VOC against 84.6% for FP32, at roughly 4% of the parameters and 2% of the operations. On NYUv2 monocular depth, it reports better error and accuracy numbers than the same FP32 baseline on every metric. The authors take these results as evidence that 1-bit networks can serve generalized dense prediction rather than only classification.","pith_inferences":["The entropy-maximization story suggests DAB would transfer to 2-bit and 4-bit quantization, where threshold placement also controls information retention; the paper does not test this.","Since CFB runs a full-precision copy alongside every binary convolution, the real-world speedup depends on how cheaply that path can be fused into a hardware pipeline; an actual latency and energy profile would settle whether the reported operation counts translate into wall-clock savings.","A natural next experiment is to apply BiDense to other dense tasks like keypoint detection or optical flow, where fine-grained localization should magnify the benefit of retaining full-precision detail.","The 'comparable to full precision' phrasing leans on the randomly initialized FP32 baseline; against the ImageNet-pretrained baseline on ADE20K the gap is larger, so the strength of the headline claim depends on which baseline is meant."],"forward_implications":["Binary dense prediction reaches within a few accuracy points of full precision on the reported benchmarks, suggesting 1-bit networks are a practical option for on-device segmentation and depth estimation.","DAB's input-dependent thresholds should keep working on unseen image distributions, since the threshold and scale track the statistics of whatever input arrives.","CFB applies to arbitrary channel-size changes, so the method transfers to any encoder-decoder or transformer-CNN hybrid beyond the ConvNeXt-UPerNet and DPT architectures tested.","The reported gains on a regression task (depth) as well as classification (segmentation) indicate that binarization need not be limited to label prediction."],"supporting_citations":[{"why":"defines canonical binary neural networks and the OPs/Params counting convention used for all efficiency numbers.","marker":"[23]"},{"why":"provides the base binary convolutional layer (sign approximation, scaled binary weights) that BiDense builds on and serves as a main baseline.","marker":"[30]"},{"why":"introduces full-precision bypasses for binary convolutions restricted to half or double channel changes, which CFB generalizes.","marker":"[3]"},{"why":"supplies the scaling-factor idea for binarized activations that DAB turns into an input-adaptive scale.","marker":"[34]"},{"why":"supplies the entropy-maximization rationale and mean-based threshold that DAB adapts to each input distribution.","marker":"[39]"},{"why":"is the ConvNeXt backbone used for the full-precision baseline that BiDense binarizes.","marker":"[33]"},{"why":"defines the dense-prediction task setups, decoder/head choices, and evaluation protocols used in the experiments.","marker":"[43]"},{"why":"is an adaptive binary baseline compared across segmentation and depth tasks.","marker":"[52]"}],"fun_headline_variants":["1-bit network matches FP32 on dense tasks at 2% compute","BiDense: binary net rivals full-precision segmentation","Adaptive binarization closes gap on dense prediction","Tiny binary model nears FP32 accuracy in vision tasks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The efficiency claim assumes that carrying a full-precision copy of every activation through every binary convolution costs almost nothing in memory and runtime, but the paper reports no latency, energy, or peak-memory measurements.","fun_headline_variants_meta":{"raw":{"variants":["1-bit network matches FP32 on dense tasks at 2% compute","BiDense: binary net rivals full-precision segmentation","Adaptive binarization closes gap on dense prediction","Tiny binary model nears FP32 accuracy in vision tasks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000179,"raw_usage":{"total_tokens":1356,"prompt_tokens":1054,"completion_tokens":302,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":670,"completion_tokens_details":{"reasoning_tokens":232}},"tokens_in":670,"tokens_out":302,"duration_ms":3292,"temperature":1.0,"reasoning_tokens":232,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:43:16.031195+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run BiDense on a mobile or edge GPU and measure end-to-end latency, peak memory, and energy against the FP32 baseline and a BiDense variant with the CFB removed. If the full-precision bypasses make runtime or memory approach the FP32 model, the paper's core efficiency claim fails.","supporting_citations":[{"cited_title":"Binarized neural networks","cited_arxiv_id":null,"evidence_quote":"defines canonical binary neural networks and the OPs/Params counting convention used for all efficiency numbers."},{"cited_title":"Reactnet: Towards precise binary neural net- work with generalized activation functions","cited_arxiv_id":null,"evidence_quote":"provides the base binary convolutional layer (sign approximation, scaled binary weights) that BiDense builds on and serves as a main baseline."},{"cited_title":"Binarized spectral compressive imaging","cited_arxiv_id":null,"evidence_quote":"introduces full-precision bypasses for binary convolutions restricted to half or double channel changes, which CFB generalizes."},{"cited_title":"Bit: Robustly binarized multi-distilled transformer","cited_arxiv_id":null,"evidence_quote":"supplies the scaling-factor idea for binarized activations that DAB turns into an input-adaptive scale."},{"cited_title":"Forward and backward information retention for accurate binary neural networks","cited_arxiv_id":null,"evidence_quote":"supplies the entropy-maximization rationale and mean-based threshold that DAB adapts to each input distribution."},{"cited_title":"A convnet for the 2020s","cited_arxiv_id":null,"evidence_quote":"is the ConvNeXt backbone used for the full-precision baseline that BiDense binarizes."},{"cited_title":"Vi- sion transformers for dense prediction","cited_arxiv_id":null,"evidence_quote":"defines the dense-prediction task setups, decoder/head choices, and evaluation protocols used in the experiments."},{"cited_title":"Adabin: Improving binary neural networks with adaptive bi- nary sets","cited_arxiv_id":null,"evidence_quote":"is an adaptive binary baseline compared across segmentation and depth tasks."}],"review_version":1}