{"id":"21bf896d-56c7-4791-a5dd-17b9e217db07","arxiv_id":"2412.13273","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A mobile-optimized optical flow network hits about 25 FPS on an iPhone 8 at 512x512 while staying competitive on KITTI and Sintel.","lead":"CompactFlowNet is an optical-flow neural network redesigned to run fast and use little memory on phones, reporting real-time speed on an iPhone 8. It combines a mobile-friendly backbone, a streamlined flow estimator, and teacher-student training to keep accuracy close to larger lightweight models.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's own test tables contradict the abstract's 'superior or comparable performance' claim: CompactFlowNet trails FDFlowNet on Sintel Clean, Sintel Final, and KITTI test.","rationale":"We agree with the reader's conditional assessment and their observation that the accuracy comparison is overstated on KITTI. However, we believe the stronger load-bearing concern is the internal contradiction between the abstract's 'superior or comparable performance' and the test-set rows in Tables 4 and 5. This is not a reproducibility gap but a direct conflict within the paper itself: CompactFlowNet is reported worse than FDFlowNet on all three public test metrics, with KITTI f1-all 14.64 vs 9.38. The reader's conversion-fairness concern is valid and would affect the speed claim, but it is speculative; the accuracy contradiction is already documented. If the authors revise the abstract to accurately describe the accuracy trade-off and release artifacts to verify speed, the paper could be conditionally accepted. Therefore the reader's CONDITIONAL verdict remains unchanged.","tokens_in":13939,"tokens_out":9452,"duration_ms":83767,"concrete_test":"As an analytical check, compute the relative errors from Tables 4 and 5: (CompactFlowNet − FDFlowNet)/FDFlowNet for Sintel Clean, Sintel Final, and KITTI f1-all. If the signs are positive on all three (as printed) and the KITTI gap is >50%, the abstract's 'superior or comparable' assertion is contradicted. To settle empirically, release the trained CompactFlowNet and submit it to the official Sintel and KITTI benchmark servers; the returned EPE/f1-all should reproduce the paper's test rows, and a comparison against FDFlowNet's leaderboard entries will decide whether the accuracy claim holds under any charitable reading.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim has two parts: real-time mobile speed and 'superior or comparable performance' to lightweight SOTA on KITTI and Sintel (abstract, Section 1). The test-set rows in Tables 4 and 5 directly refute the accuracy part. CompactFlowNet's reported test errors are: Sintel Clean 4.43 vs FDFlowNet 3.71; Sintel Final 5.55 vs 5.11; KITTI 2015 f1-all 14.64 vs 9.38. On every public test benchmark, CompactFlowNet is worse than FDFlowNet, with the KITTI gap exceeding 50% relative error. The paper only emphasizes parenthesized train-set numbers where the differences are small or favorable, but benchmarks conventionally judge on test. The Discussion's statement that CompactFlowNet 'aligns closely with' FDFlowNet is inconsistent with these numbers. Because the abstract explicitly promises 'superior or comparable performance,' the paper's own results falsify a core part of the central claim. This is an internal inconsistency, not a matter of missing artifacts or conversion fairness; it stands regardless of speed. The speed advantage (40 ms iPhone 8) may be genuine, but the overall claim as written is overstated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"CompactFlowNet is a proposed mobile-optimized optical flow network built on a PWC-Net-style architecture with a MobileNetV3 backbone, a sequentially connected flow estimator that removes dense concatenation, depthwise separable convolutions, and a reduced-depth flow refiner. The paper reports 40 ms average latency on an iPhone 8 at 512x512 input, lower peak memory than several lightweight competitors, and claims 'superior or comparable performance' to state-of-the-art lightweight models on KITTI and Sintel. Training uses a distillation pipeline with the authors' own heavy PWC-Net+MobileNetV3 model as teacher, and results are reported on Sintel and KITTI benchmarks as well as on-device latency and memory across four iPhone models.","tokens_in":14173,"tokens_out":5272,"duration_ms":46360,"significance":"If the reported speed and memory numbers are reproducible, the paper makes a useful engineering contribution: it demonstrates that a full optical flow network can run at interactive rates on a 2017-era phone, and it provides a cross-device latency and memory comparison for several lightweight methods. The paper ships no code or weights, however, and the accuracy claim is contradicted by the paper's own public test-set numbers, which show CompactFlowNet trailing FDFlowNet on every benchmark. The speed advantage may be real, but the overall claim as stated is overstated and the benchmarking methodology is not sufficiently specified to assess it.","major_comments":[{"comment":"The abstract and Section 1 claim 'superior or comparable performance' to state-of-the-art lightweight models, but the paper's own public test-set results contradict this. On Sintel Clean test, CompactFlowNet reports 4.43 AEPE versus FDFlowNet's 3.71; on Sintel Final test, 5.55 versus 5.11; and on KITTI 2015 test (f1-all), 14.64 versus 9.38. CompactFlowNet is worse than FDFlowNet on every public test benchmark, with a relative KITTI gap exceeding 50%. The Discussion's statement (Section 4.3) that CompactFlowNet 'aligns closely with' FDFlowNet is inconsistent with these numbers, which are the conventional basis for benchmark comparison. The claim must be revised or re-scoped, for example to 'competitive accuracy at a fraction of the latency,' and the Discussion should honestly state the accuracy trade-off.","section":"Abstract; Tables 4 and 5; Section 4.3"},{"comment":"The on-device speed and memory comparison is under-specified. The text says all models were 'implemented in PyTorch and converted to TFLite,' but it does not report the conversion tool versions, whether weights were quantized or cast to float16, which TFLite delegate (GPU or CPU) was used, or any per-model conversion settings. If competitor models convert less favorably, for example due to unsupported operators falling back to CPU, the claimed speed and memory advantage of CompactFlowNet could be an artifact of the conversion pipeline rather than of the architecture. The paper also gives no variance or error bars for the latency measurements, despite averaging over 100 runs. This is load-bearing for the paper's central speed claim.","section":"Section 4.2; Tables 6 and 7"},{"comment":"The paper provides no code, pretrained weights, or complete training configuration, and it refers to 'specific augmentation parameters in Supplementary Materials' that are not present in the arXiv submission. The training pipeline includes many undocumented choices, such as the exact OneCycle schedule details, batch size, augmentation ranges, and distillation teacher configuration, which would be needed to reproduce the models. For a paper whose contribution is an architecture plus a training recipe, this lack of reproducibility evidence weakens the verification of both the accuracy and the speed claims.","section":"Section 4.1; Section 3.1"}],"minor_comments":[{"comment":"The abstract states 'the first real-time mobile neural network for optical flow prediction,' while the introduction uses 'To our knowledge, this is the first compact and memory-efficient model...' These are different claims; the hedging should be consistent, and the novelty claim should be supported by a more thorough survey of mobile deployment efforts.","section":"Section 1"},{"comment":"The parenthesized values in Table 3 are not sufficiently explained; the caption says they 'represent the results of networks that were trained on the same dataset,' but the reader has to infer which dataset is meant. Please clarify in the caption.","section":"Table 3"},{"comment":"The bold font for real-time compatible latency is helpful, but the threshold for real-time, for example 25 FPS or 40 ms, is not defined in the text or caption; please state it explicitly.","section":"Table 6"},{"comment":"The description of depthwise separable convolutions is standard, but it would help to state whether they replace all standard convolutions in the flow estimator and refiner, or only some; the text says 'adopted in both the flow estimator and flow refiner blocks' without specifying the extent.","section":"Section 3.2"},{"comment":"Reference [54] is cited for GPU latency and memory of RAFT and PWC-Net, but the numbers given in the text, such as 107 ms for RAFT, should be attributed explicitly to that paper to avoid ambiguity.","section":"Section 2"}],"recommendation":"major_revision","confidential_remarks":"The core engineering result, real-time optical flow on an iPhone 8, is potentially interesting and within the journal's scope, but the accuracy claim is internally falsified and the benchmarking protocol lacks essential details. If the authors supply code and weights, provide a detailed conversion specification, and honestly re-frame the accuracy comparison as a trade-off rather than superiority, a revised version could be acceptable. The 'first' novelty claim also needs a more careful literature check before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a real engineering contribution with an overstated accuracy claim. The speed and memory numbers on iPhone 8 are impressive, and the architecture changes are sensible. But the abstract promises 'superior or comparable performance' to lightweight SOTA, and the paper's own test tables contradict that: on Sintel Clean test CompactFlowNet is 4.43 vs FDFlowNet 3.71, on Sintel Final 5.55 vs 5.11, and on KITTI test 14.64 vs 9.38 f1-all. That's not a minor wording issue; it's a load-bearing claim that fails against the paper's own evidence.\n\nWhat's genuinely new: a mobile-optimized PWC-Net variant with a sequentially connected flow estimator (no dense concatenation), depthwise separable convolutions in the estimator and refiner, a MobileNetV3 backbone, and a distillation pipeline from a heavier teacher. The on-device latency comparison across iPhone 8/XR/12/14 Pro and three input resolutions is useful data for the community; to my knowledge this is the first measurement of these lightweight flow models on phones. The ablation shows the pretrained backbone and distillation each give a clear gain, which is credible.\n\nThe soft spots are real but addressable. First, the accuracy overclaim: the paper emphasizes parenthesized train-set numbers where CompactFlowNet looks good, but benchmarks are conventionally judged on test, and the test numbers are clearly worse than FDFlowNet, with the KITTI gap over 50% relative error. Claiming 'aligns closely' is unfair to the reader. Second, the benchmarking protocol is under-specified: no code or weights, no TFLite conversion precision or delegate details, and no error bars on latency. The conversion from PyTorch to ONNX to TF to TFLite can easily privilege one architecture over another depending on operator support and kernel selection, so the speed comparison isn't fully checkable without those details. The missing supplementary materials on the arXiv version don't help.\n\nWho is this for: people working on efficient optical flow, mobile computer vision, and model deployment. With artifact release and a corrected accuracy claim it would be a solid paper; without them, the speed result is plausible but not verifiable.\n\nRecommendation: yes, send it to peer review. The engineering result is likely real and the mobile benchmarking is valuable. A serious referee can push for the claims to be matched to the data and for the artifacts to be released. Desk rejection would be too harsh; the paper just needs an honest rewrite and more transparency.","headline":"A useful mobile optical flow system with a real speed/memory win, but the abstract's accuracy claim is contradicted by the paper's own test tables.","tokens_in":14733,"tokens_out":2771,"would_cite":false,"duration_ms":25546,"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":"CompactFlowNet runs in 40 ms per 512x512 frame on an iPhone 8 while keeping accuracy on par with lightweight baselines.","keywords":["optical flow estimation","mobile inference","real-time video processing","CompactFlowNet","knowledge distillation","depthwise separable convolutions","PWC-Net"],"falsifier":"Reproduce the benchmark with the models converted under identical settings (same precision, same delegate, same iPhone OS) and check the 40 ms latency and 118 MB peak memory at 512x512 on an iPhone 8; if the margin over FastFlowNet disappears or the memory numbers grow past the reported values, the central claim fails.","tokens_in":13728,"feed_emoji":"📱","tokens_out":9316,"duration_ms":78463,"temperature":0.7,"pith_summary":"CompactFlowNet is an optical flow network built to run on phones. Optical flow predicts the per-pixel motion between two frames, and it is the backbone of video stabilization, tracking, action recognition, and video generation, so a fast on-device version makes those tasks possible without uploading video to a server. The paper's central claim is that CompactFlowNet is the first model to combine real-time mobile inference with accuracy competitive against lightweight baselines: 40 ms per 512x512 frame on an iPhone 8, 13 ms on an iPhone 14 Pro, and a peak memory footprint of 118 MB at that resolution. The design starts from a PWC-Net-style pyramid and reworks the expensive flow estimator, the refiner, and the backbone, then closes the accuracy gap with knowledge distillation from a heavier teacher.","feed_headline":"Optical flow hits 40 ms on an iPhone 8","feed_subtitle":"A compact model delivers real-time motion estimation on phones with accuracy close to lightweight baselines.","key_machinery":"The load-bearing mechanism is architectural: the flow estimator, which dominates parameter count, FLOPs, and latency, is changed from a densely connected block to a sequentially connected block. Dense connections force every layer to receive concatenated maps from all earlier layers, keeping large tensors alive in memory and inflating both latency and the downstream refiner's input size; sequential connections remove that storage and compute burden. The other components are depthwise separable convolutions in the estimator and refiner, a shortened refiner with channels 128, 64, 32, and 2, and a MobileNetV3 backbone chosen by an on-device comparison of candidate backbones at 512x512 on an iPhone 8. Training adds a second mechanism: knowledge distillation from the heavier MobileNetV3-backed PWC-Net teacher, with loss $L = L_{\\mathrm{sup}} + \\gamma L_{\\mathrm{dist}}$ at $\\gamma = 0.1$, combined with an Autoflow pre-training dataset, extra augmentations, a OneCycle learning rate, and gradient clipping.","core_discovery":"The paper argues that a PWC-Net-style coarse-to-fine optical flow model can be made mobile-friendly without sacrificing accuracy, by attacking the flow estimator block. Profiling on the iPhone 8 shows the estimator accounts for roughly 6.05 of the 8.75 million parameters and 131 of the 212 ms total latency of the original PWC-Net at 512x512. The authors replace the estimator's dense concatenation with a sequential connection pattern, adopt depthwise separable convolutions in the estimator and refiner, shrink the refiner from seven layers to four with channels 128, 64, 32, and 2, and swap the backbone for MobileNetV3 while keeping the full model under five million parameters. A heavy teacher model, the original architecture equipped with the MobileNetV3 backbone, is then used to distill CompactFlowNet, with a per-pixel L2 supervision loss plus a distillation term weighted at 0.1. The resulting model reaches 40 ms latency on an iPhone 8, 26 ms on an iPhone XR, 19 ms on an iPhone 12, and 13 ms on an iPhone 14 Pro at 512x512, with peak memory of 118 MB at that resolution; it also beats FastFlowNet and SpyNet on the Sintel test sets and runs faster than all compared lightweight models across three resolutions and four devices.","pith_inferences":["The profiling table implies that dense feature concatenation, not the backbone, is the main latency and memory bottleneck in lightweight flow models; a direct activation-memory profiler could verify this on other PWC variants.","The accuracy gain from distillation (Sintel Final validation from 4.05 to 3.36, KITTI 2015 train from 2.14 to 1.91) suggests that a stronger teacher from a different model family might reduce the remaining error, although the paper reports trying such teachers without improvement.","The KITTI 2015 test f1-all of 14.64 is higher than FDFlowNet's 9.38, so the 'superior or comparable' claim should be read as benchmark-dependent; the speed and memory story does not depend on that comparison."],"forward_implications":["Tasks that depend on optical flow, such as stabilization, tracking, restoration, frame interpolation, and generative video editing, can be executed on-device without uploading footage.","The model runs at real-time speed (40 ms, 25 FPS) on an iPhone 8 and 13 ms on an iPhone 14 Pro at 512x512, with peak memory of 118 MB.","At Full HD (1080x1920) input it reaches nearly 10 FPS on an iPhone 14 Pro, and outperforms FastFlowNet and SpyNet on the Sintel test sets.","The under-10 MB float16 weight size and reduced memory footprint make the model compatible with the constraints the paper cites for mobile deployment."],"supporting_citations":[{"why":"Supplies the PWC-Net pyramid-warping-cost-volume architecture that CompactFlowNet starts from and reworks.","marker":"[51]"},{"why":"Supplies the MobileNetV3 backbone that replaces the original feature extractor and improves accuracy without slowing inference.","marker":"[17]"},{"why":"Supplies the Autoflow pre-training dataset the paper credits with improving accuracy for PWC-Net and RAFT.","marker":"[53]"},{"why":"Supplies the OneCycle learning-rate and gradient-clipping training recipe adopted for the compact model.","marker":"[52]"},{"why":"Supplies FastFlowNet, the main lightweight speed competitor CompactFlowNet is compared against.","marker":"[30]"},{"why":"Supplies the enhanced PWC-Net+ training pipeline and the GPU latency and memory comparison numbers for heavier models.","marker":"[54]"},{"why":"Supplies the depthwise separable convolution building block used in the redesigned estimator and refiner.","marker":"[18]"}],"fun_headline_variants":["CompactFlowNet: real-time optical flow on mobile","Optical flow hits 40ms on iPhone 8, no server needed","Mobile optical flow: 40ms inference on iPhone 8","First real-time mobile optical flow net","Phone-grade real-time optical flow at 40ms"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The speed and memory comparisons assume that every model is converted to the mobile format with equally fair settings, so the measured differences on the phone reflect the architectures rather than conversion or precision artifacts.","fun_headline_variants_meta":{"raw":{"variants":["CompactFlowNet: real-time optical flow on mobile","Optical flow hits 40ms on iPhone 8, no server needed","Mobile optical flow: 40ms inference on iPhone 8","First real-time mobile optical flow net","Phone-grade real-time optical flow at 40ms"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000904,"raw_usage":{"total_tokens":3937,"prompt_tokens":1040,"completion_tokens":2897,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":656,"completion_tokens_details":{"reasoning_tokens":2817}},"tokens_in":656,"tokens_out":2897,"duration_ms":22942,"temperature":1.0,"reasoning_tokens":2817,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T13:17:30.780565+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reproduce the benchmark with the models converted under identical settings (same precision, same delegate, same iPhone OS) and check the 40 ms latency and 118 MB peak memory at 512x512 on an iPhone 8; if the margin over FastFlowNet disappears or the memory numbers grow past the reported values, the central claim fails.","supporting_citations":[{"cited_title":"Searching for mo- bilenetv3","cited_arxiv_id":null,"evidence_quote":"Supplies the MobileNetV3 backbone that replaces the original feature extractor and improves accuracy without slowing inference."},{"cited_title":"Autoflow: Learning a better training set for optical flow","cited_arxiv_id":null,"evidence_quote":"Supplies the Autoflow pre-training dataset the paper credits with improving accuracy for PWC-Net and RAFT."},{"cited_title":"Models matter, so does training: An empirical study of cnns for optical flow estimation","cited_arxiv_id":null,"evidence_quote":"Supplies the OneCycle learning-rate and gradient-clipping training recipe adopted for the compact model."},{"cited_title":"Fastflownet: A lightweight network for fast optical flow estimation","cited_arxiv_id":null,"evidence_quote":"Supplies FastFlowNet, the main lightweight speed competitor CompactFlowNet is compared against."},{"cited_title":"Disentan- gling architecture and training for optical flow","cited_arxiv_id":null,"evidence_quote":"Supplies the enhanced PWC-Net+ training pipeline and the GPU latency and memory comparison numbers for heavier models."}],"review_version":1}