{"id":"9dbe8ca3-db74-46ea-9651-77a4a706a893","arxiv_id":"2412.12672","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A new filter pruning method uses graph-based selection on spatial feature maps to improve the accuracy-efficiency trade-off of compressed segmentation networks.","lead":"This paper introduces SIRFP, a way to compress semantic segmentation neural networks by removing redundant channels while keeping their location-sensitive information. It matters because smaller, faster segmentation models are what make real-time applications like self-driving cars and mobile photo editing practical.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1's unpruned baseline (81.6 mIoU) contradicts Table 7 and Appendix C (79.3 mIoU) for the same Deeplabv3-ResNet50 on Cityscapes, making the reported pruning margins hard to interpret until reconciled.","rationale":"I considered the EMA-staleness concern raised by the reader. It is real, but it is partly mitigated by the alpha=0.99 decay: after 4000 pre-training iterations the contribution of the first iterations is roughly 10^-17, so only the last about 100 iterations materially shape the edge weights. The baseline inconsistency, however, is not mitigated and is directly visible in the manuscript. The numerical anchor of every headline comparison is the unpruned model; if that anchor shifts by 2.3 mIoU, the claimed advantages over FPGM and DCFP cannot be interpreted. The reader did flag the baseline discrepancy in their rationale, but their weakest_assumption was the EMA accumulation; I see the baseline discrepancy as more load-bearing, hence 'partial'. A no-pruning fine-tuning control is a cheap, decisive experiment. The verdict remains conditional: the method may still be good, but the published numbers need reconciliation before the trade-off claim is accepted.","tokens_in":17747,"tokens_out":5526,"duration_ms":48693,"concrete_test":"Run Algorithm 2 on Deeplabv3-ResNet50/Cityscapes with zero channels pruned: pre-train with the Table 1 recipe (including GSRL), then apply progressive 'pruning' steps and fine-tuning for exactly the same number of iterations as the pruned models, and report mIoU. If this no-pruning control reaches roughly 81.3, the pruning decision is not the source of the gain; recompute Table 1 margins against this control and against FPGM/DCFP executed with the same schedule. Separately, report which configuration (GSRL on/off, same eval) produces the 81.6 versus 79.3 unpruned numbers.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is the state-of-the-art parameter/mIoU trade-off in Table 1: Ours-60% reaches 81.3 versus FPGM 80.2 and DCFP 80.9. However, the paper reports two incompatible unpruned baselines for the same model and dataset. Table 1 gives 'Unpruned 81.6', while Table 7 (the TensorRT speed comparison) and Appendix C (Tables 8 and 9, ablations of resolution and metric) give 'Unpruned 79.3'. If the true baseline is 79.3, Ours-60% (81.3) is 2.0 points above the unpruned model, which cannot be caused by removing 60% of FLOPs; it must reflect extra fine-tuning, GSRL, progressive pruning, or evaluation differences. Then the claimed margins over FPGM and DCFP are not attributable to the spatial-aware redundancy metric unless those baselines are re-run under the identical pipeline. The paper does not include a no-pruning control that receives the same fine-tuning schedule as pruned models, nor does it reconcile the two baseline numbers. This is the load-bearing weakness: the empirical case for the method depends on a comparison whose reference point is ambiguous.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SIRFP, a structural pruning method for semantic segmentation. The key ideas are: (i) measuring channel redundancy with a spatial-aware JS-divergence metric computed on full-resolution feature maps, accumulated via EMA during training; (ii) formulating the selection of retained channels as a Maximum Edge Weight Clique Problem (MEWCP); and (iii) solving it with a greedy heuristic (EHGP). Experiments on Cityscapes, ADE20K, COCO-stuff, COCO object detection, and ImageNet report favorable parameter/FLOPs versus accuracy trade-offs, including a claimed best parameter-mIoU trade-off on Cityscapes relative to pruning and real-time segmentation methods.","tokens_in":17854,"tokens_out":7343,"duration_ms":65618,"significance":"If the empirical claims hold, the spatial-aware redundancy formulation is a meaningful step for pruning location-sensitive models, and the greedy MEWCP solver is a practical contribution. The paper includes useful ablations (feature resolution, redundancy metric, solver runtime), reports results across multiple datasets and tasks, and releases code. The main concern is that the central Cityscapes comparison rests on inconsistent unpruned-baseline numbers, and several algorithmic details are stated ambiguously. These issues are fixable but currently prevent the results from being interpreted reliably.","major_comments":[{"comment":"The unpruned Deeplabv3-ResNet50 baseline on Cityscapes is reported as 81.6 mIoU in Table 1 but as 79.3 mIoU in Table 7 and in Appendix C, Tables 8 and 9. The central claim that Ours-60% (81.3) outperforms FPGM (80.2) and DCFP (80.9) is made against the 81.6 baseline. If the correct baseline is 79.3, then Ours-60% is 2.0 points above the unpruned model, which cannot be explained by pruning alone; it implies that extra fine-tuning, GSRL, progressive pruning, or evaluation differences contribute to the gain. The manuscript must reconcile the two baseline numbers and add a no-pruning control that receives the same fine-tuning schedule as the pruned models before the advantage over FPGM and DCFP can be attributed to the spatial-aware metric. Table 7 also reports DCFP-60% as 80.7, whereas Table 1 reports 80.9, further indicating that the comparison pipeline is not stable across tables.","section":"Table 1 vs Table 7 and Appendix C (Tables 8-9)"},{"comment":"Algorithm 1 is stated inconsistently with the surrounding text and with Appendix D. The text says the task is to delete C_l - b channels iteratively, but the loop in Algorithm 1 reads 'for each prune iteration t_p in [b]'; since b denotes the number of kept channels in Eq. (2), the loop bound should be C_l - b. In addition, Appendix D says that after collecting the sums s_k, the global threshold is set so that 'pruning all the channels in the whole network whose s_k >= s_th' satisfies the FLOPs target, whereas Algorithm 1 always removes the vertex with the minimum s_i. These two descriptions identify opposite sets of channels. The algorithm must be restated unambiguously for the method to be reproducible.","section":"Algorithm 1 and Appendix D (global threshold)"},{"comment":"Algorithm 2 accumulates the edge-weight matrix A^l via EMA starting from the first pre-training iteration, with alpha fixed at 0.99. Because early-training features are not representative of the final trained network, the running average may be dominated by stale statistics when pruning decisions are made. The paper does not ablate the EMA decay alpha or the accumulation start time, nor does it provide evidence that the running average converges to the final network's redundancy pattern. This design choice is load-bearing for the claimed benefit of the spatial-aware metric, and an ablation is needed to show that the reported results do not depend on a particular warm-up or decay value.","section":"Algorithm 2 and Eq. (4)"},{"comment":"The ImageNet comparison in Table 5 is not controlled for training schedule. Appendix D reports that SIRFP is trained for 250 epochs with a cosine schedule, pruning every 2 epochs from epoch 2 to 180, while the paper does not report the schedules used for the listed baselines, which are typically trained for fewer epochs. A longer training schedule alone can improve accuracy, so the 77.35 top-1 accuracy at 2.0G in Table 5 cannot currently be attributed to the pruning criterion. The authors should either run all methods under the same schedule or provide the training schedule used for each baseline.","section":"Table 5 and Appendix D (ImageNet schedules)"}],"minor_comments":[{"comment":"Equation (3) writes F_i^l, F_j^l as elements of R^{C_l x H x W}, but these are per-filter feature maps and should be R^{H x W} (or R^{1 x H x W}). The implementation in Appendix C states that features are normalized as probability score maps for Dice/KL/JS, but this normalization is not described in the main text and should be stated before Eq. (3).","section":"Eq. (3)"},{"comment":"The initial value of the EMA edge weight a^0_ij is not specified. If it is initialized differently (e.g., to 0 versus to the first observed value), the early updates and potentially the pruning decisions change; please specify the initialization.","section":"Eq. (4)"},{"comment":"The Limitation section says that for a well-trained network SIRFP must inference on training samples to acquire feature maps, but Algorithm 2 accumulates redundancy during pre-training. Please clarify when EMA accumulation is performed in the actual pipeline and whether the released code supports both settings.","section":"Algorithm 2 / Limitation"},{"comment":"The text repeatedly refers to 'mutual information redundancy', but the proposed metric is the Jensen-Shannon divergence between feature distributions (Eq. (3)). Please clarify the relationship between the two notions, since JS divergence is not mutual information.","section":"Abstract / Introduction"},{"comment":"The appendix contains a typo: 'Talbe 10' should be 'Table 10'. The phrase 'batch of size' should also be 'batch size'.","section":"Appendix D"}],"recommendation":"major_revision","confidential_remarks":"The inconsistency between the 81.6 mIoU unpruned baseline in Table 1 and the 79.3 mIoU baseline in Table 7 and Appendix C is serious enough that I recommend asking the authors to provide the exact training, fine-tuning, and evaluation settings for each table, as well as the no-pruning control. I do not see evidence of misconduct, but the discrepancy must be resolved before the paper's central claim can be accepted. The manuscript is within the scope of the venue and the core idea is worth pursuing, so a major revision with additional controlled experiments seems appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper does something real. Replacing pooled or weight-based similarity with a JS-divergence redundancy computed on full-resolution feature maps is a sensible task-specific move, and the greedy clique solver makes the approach computationally practical. But the main Cityscapes comparison is undercut by two incompatible unpruned baselines for the same model (81.6 in Table 1, 79.3 in Tables 7–9). That needs reconciling, along with a same-schedule no-pruning control, before the gains over DCFP can be credited to the metric.\n\nWhat is actually new: the package of un-pooled feature redundancy via JS divergence, EMA accumulation of edge weights during training, and the MEWCP formulation with an O(n^2) greedy deletion. I don't see that exact combination in the prior work cited. The ablations on feature resolution and metric choice are useful; the result that full resolution helps is consistent with the task motivation, and the explanation for why dot product underperforms is plausible.\n\nWeak spots, in order of severity. First, the baseline inconsistency is load-bearing. If the true unpruned mIoU is 79.3, then Ours-60% at 81.3 is two points above the unpruned model. Pruned models can beat a weak baseline after heavy fine-tuning, but with no control that receives the same fine-tuning schedule, GSRL, and progressive pruning, the pruning criterion is not isolated. Second, the metric and hyperparameters are selected on Cityscapes, the same benchmark used for the headline comparison, so some tuning leakage is likely. Third, the claimed equivalence between the linear pruning objective and the quadratic MEWCP is informal; eq (1) is linear in the mask, eq (2) is quadratic, and the reduction only works by fixing the mask on the right-hand side. The greedy solver is a fine heuristic, but the equivalence is a stretch. The ImageNet results use a 250-epoch recipe that may explain part of the margin over earlier numbers.\n\nThese are fixable. Reconcile the baselines, add a no-pruning control, and the method likely survives as a modest but genuine contribution. The paper is transparent about its EMA cost in the limitation section. I would send it to peer review—the task-specific motivation is sound and the novelty is real—but the baseline issue should be settled before acceptance.","headline":"Real novelty in the spatial-aware pruning metric, but the headline comparison is undermined by two inconsistent unpruned baselines that need to be reconciled.","tokens_in":18580,"tokens_out":2860,"would_cite":false,"duration_ms":25066,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T07","68T45","68R10","05C69","05C85"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that channel pruning for semantic segmentation should be driven by spatial redundancy in feature maps, not weight similarity, and that its SIRFP method achieves higher mIoU at matched FLOPs than prior pruning methods.","keywords":["filter pruning","semantic segmentation","spatial-aware redundancy","maximum edge weight clique problem","greedy channel pruning","feature-map redundancy","model compression","EMA statistics"],"falsifier":"Re-run the Deeplabv3-ResNet50/Cityscapes pipeline with the edge-weight matrix accumulation started only after the first half of pre-training, or with the EMA hyperparameter alpha set to 0.9 instead of 0.99. If the 60%-FLOPs mIoU stays near 81.3 and the margin over FPGM and DCFP persists, the EMA-schedule assumption is not load-bearing; if the mIoU drops by more than about 0.5 points, the reported results depend on statistics gathered from an immature network.","tokens_in":17356,"feed_emoji":"✂️","tokens_out":10375,"duration_ms":82703,"temperature":0.7,"pith_summary":"Most filter-pruning methods are built for image classification and score channels by weight similarity or pooled features, which discards where in the image a channel fires. This paper argues that semantic segmentation needs a spatial-aware redundancy measure, and proposes SIRFP, which compares full-resolution feature maps with the Jensen-Shannon divergence and prunes the channels whose spatial patterns are most duplicated. The paper claims that selecting kept channels to minimize pairwise redundancy is exactly a maximum-edge-weight clique problem, and that a greedy one-by-one deletion solver makes this practical even for layers with more than 1000 channels. If correct, the method means location-sensitive tasks need their own pruning metrics, and the reported results show higher mIoU at 60–80% FLOPs reduction than FPGM and DCFP on Cityscapes, with similar gains on ADE20K and COCO-Stuff.","feed_headline":"Prune by spatial redundancy to keep segmentation accuracy","feed_subtitle":"SIRFP's feature-map metric beats weight-based pruning at matched FLOPs on Cityscapes.","key_machinery":"The engine is a complete edge-weighted graph $G=(V,E)$ whose vertices are the channels of a layer and whose edge weight $a_{ij}$ is the EMA-accumulated value of $1 - r_{ij}$, where $r_{ij}$ is the Jensen-Shannon divergence between the two channels' full-resolution feature maps (Equation 3). Pruning becomes the Maximum Edge Weight Clique Problem: choose the $b$ vertices whose induced subgraph has the largest total edge weight, which is equivalent to keeping the least mutually redundant channels. The efficient heuristic greedy pruning (EHGP) then repeatedly removes the vertex with the smallest current sum of incident edge weights, updating the remaining sums in $O(n^2)$ total time, and a global FLOPs threshold converts per-layer sums into a pruning decision across the network.","core_discovery":"The central claim is that a channel is redundant not when its weights resemble another channel's weights, but when its output feature map, taken at full spatial resolution, carries information already present in other kept channels. SIRFP encodes this as edge weights $a_{ij}$ between every pair of channels, set to the EMA-accumulated complement of a Jensen-Shannon divergence between their un-pooled feature maps, and keeps the budget of channels whose total edge weight is maximal. The paper shows that this objective is the Maximum Edge Weight Clique Problem, that the one-vertex-deletion case has a closed-form solution, and that iterating that deletion gives a greedy solver with $O(n^2)$ average complexity. On Deeplabv3-ResNet50 with Cityscapes, the method reaches 81.3% mIoU at 60% FLOPs reduction, versus 80.2% for FPGM and 80.9% for DCFP, and the margin widens at higher compression ratios; the same pipeline also transfers to ADE20K, COCO-Stuff, object detection, and image classification.","pith_inferences":["An untested extension is that the same spatial-redundancy objective should help other dense prediction tasks, such as monocular depth estimation or instance segmentation, where localization errors are directly penalized; running SIRFP on those tasks would separate the metric's contribution from the MEWCP solver's.","Their own runtime table shows the exact MEWCP solver reaches 81.4% mIoU versus 81.3% for the greedy EHGP, so a stronger local-search heuristic that looks two deletions ahead might capture most of the remaining 0.1-point gap without the solver's 10,800-second cost.","The paper fixes EMA alpha at 0.99 and starts accumulation at iteration 1 without ablation; a testable prediction is that beginning accumulation later, or using a smaller alpha, would shift which channels survive and either shrink or widen the gap to DCFP.","If feature-map redundancy patterns stabilize early in training, the method could be applied to an already-trained network with only a short feature-collection pass, which would remove the paper's own stated limitation that it must run during training."],"forward_implications":["At 60% FLOPs reduction on Deeplabv3-ResNet50 over Cityscapes, SIRFP yields 81.3% mIoU, 1.1 points above FPGM and 0.4 above DCFP with slightly fewer remaining FLOPs.","At 70% and 80% reductions the margins over DCFP grow to 1.1 and 0.6 mIoU, showing the spatial metric degrades more gracefully under aggressive compression.","The same EMA-accumulated edge weights transfer across tasks: 44.9% mIoU on ADE20K and 34.8% on COCO-Stuff at 60% reduction, 26.6 AP on SSD-ResNet50/COCO at 50% FLOPs cut, and 77.35% top-1 on ImageNet at 2.0G FLOPs.","Because redundancy statistics are accumulated during training via EMA, the pipeline avoids the separate feature-collection inference pass used by HRank- and CHIP-style pruning methods."],"supporting_citations":[{"why":"Defines FPGM, the weight-similarity pruning baseline that SIRFP claims is spatial-agnostic and must beat (80.2 vs 81.3 mIoU at 60% FLOPs cut).","marker":"He et al. 2019"},{"why":"Defines DCFP, the main EMA-based segmentation pruning competitor, and supplies the GSRL loss used in all segmentation experiments and the 80.9/79.8/78.8 mIoU numbers SIRFP is compared against.","marker":"Wang et al. 2024"},{"why":"Introduces CHIP, a data-driven feature-based pruning method that uses pooled features; it is the key contrast case for why un-pooled spatial features matter.","marker":"Sui et al. 2021"},{"why":"Introduces HRank, another data-driven feature-based pruning method whose pooled-vector importance scores are the spatial-agnostic approach SIRFP argues is insufficient.","marker":"Lin et al. 2020a"},{"why":"Formulates the maximum edge weight clique problem and gives the explicit solution for deleting a single vertex, which the greedy EHGP iterates.","marker":"Späth 1985"},{"why":"Establishes the NP-hardness of MEWCP, justifying the need for the efficient greedy solver.","marker":"Hosseinian et al. 2017"},{"why":"Provides the exact MEWCP solver whose 10,800-second runtime and 81.4 mIoU are compared with EHGP's 285 seconds in Table 6.","marker":"Hosseinian, Fontes, and Butenko 2020"},{"why":"Provides the Cityscapes dataset, the benchmark for the paper's headline parameter-mIoU trade-off results.","marker":"Cordts et al. 2016"}],"fun_headline_variants":["Prune by spatial redundancy to keep segmentation sharp","Spatial-aware filter pruning beats weight-based on Cityscapes","Graph-based pruning with feature-map redundancy for segmentation","SIRFP: prune channels by spatial info, not weights"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the running average of channel-redundancy scores, collected from the network while it is still being trained, accurately reflects which channels are redundant in the fully trained network; if early-training features are uninformative or the feature distribution shifts substantially, the pruning decisions would be based on stale statistics.","fun_headline_variants_meta":{"raw":{"variants":["Prune by spatial redundancy to keep segmentation sharp","Spatial-aware filter pruning beats weight-based on Cityscapes","Graph-based pruning with feature-map redundancy for segmentation","SIRFP: prune channels by spatial info, not weights"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000428,"raw_usage":{"total_tokens":2216,"prompt_tokens":1000,"completion_tokens":1216,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":616,"completion_tokens_details":{"reasoning_tokens":1151}},"tokens_in":616,"tokens_out":1216,"duration_ms":9688,"temperature":1.0,"reasoning_tokens":1151,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T13:51:08.299052+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the Deeplabv3-ResNet50/Cityscapes pipeline with the edge-weight matrix accumulation started only after the first half of pre-training, or with the EMA hyperparameter alpha set to 0.9 instead of 0.99. If the 60%-FLOPs mIoU stays near 81.3 and the margin over FPGM and DCFP persists, the EMA-schedule assumption is not load-bearing; if the mIoU drops by more than about 0.5 points, the reported results depend on statistics gathered from an immature network.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines DCFP, the main EMA-based segmentation pruning competitor, and supplies the GSRL loss used in all segmentation experiments and the 80.9/79.8/78.8 mIoU numbers SIRFP is compared against."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Introduces CHIP, a data-driven feature-based pruning method that uses pooled features; it is the key contrast case for why un-pooled spatial features matter."},{"cited_title":"B.; Butenko, S.; Nardelli, M","cited_arxiv_id":null,"evidence_quote":"Establishes the NP-hardness of MEWCP, justifying the need for the efficient greedy solver."},{"cited_title":"B.; and Butenko, S","cited_arxiv_id":null,"evidence_quote":"Provides the exact MEWCP solver whose 10,800-second runtime and 81.4 mIoU are compared with EHGP's 285 seconds in Table 6."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the Cityscapes dataset, the benchmark for the paper's headline parameter-mIoU trade-off results."}],"review_version":1}