{"id":"940bebbe-e519-4793-a0ec-c086d87bed7b","arxiv_id":"1908.06692","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":6,"one_line_summary":"Auxiliary video losses help an under-trained OSVOS on DAVIS-2016, but the gains are small and the comparison setting is non-standard.","lead":"This paper proposes adding auxiliary video-loss heads, including a video-id classifier and metric-learning losses, to the parent-network training step of OSVOS for one-shot video object segmentation. The authors report small J-mean gains over OSVOS on DAVIS-2016, but only at a non-standard 10k-iteration setting where the OSVOS baseline is below its published result.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Central claim rests on 10k-iteration comparison against an un-converged OSVOS baseline; no video-loss results at the 20k setting that the paper itself says replicates OSVOS.","rationale":"The reader's weakest assumption identifies the same decisive problem: the headline improvement is measured at 10k iterations, while the paper itself says OSVOS needs about 20k to reproduce the original numbers. The paper reports no OSVOS-VL result at 20k, so the central claim that video loss 'can improve the segmentation performance significantly' is not established at the standard setting. I agree with the REJECT verdict for that reason. I also note the single-run nature and large per-sequence regressions, which reinforce the concern but are secondary to the missing 20k comparison. The proposed fix is straightforward: collect 20k results for the proposed method and compare to the 20k OSVOS baseline. If the improvement disappears, the claim is not supported; if it survives, the reader's rejection should be reconsidered. Since that evidence is absent, no verdict change is warranted.","tokens_in":8195,"tokens_out":5936,"duration_ms":64892,"concrete_test":"Re-run the Table 1 and Table 2 experiments (both backbones, all four loss variants) with online fine-tuning extended to 20k iterations under the same protocol, using at least three seeds and reporting per-sequence J-mean. Then compare OSVOS-VL at 20k against OSVOS at 20k. If the margins remain similar to the 10k margins, the protocol concern is resolved. If the margins shrink below roughly 1 J-mean or reverse, the central claim should be reframed as a convergence-speed improvement rather than a final-performance improvement.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.2 states that all experiments use 10k online fine-tuning iterations and that at roughly 20k iterations the OSVOS baseline replicates the numbers originally reported in [2]. The Table 1 OSVOS values (75.0 VGG16, 65.7 MobileNet) are therefore 10k checkpoints, not the converged OSVOS the abstract's comparison implies. The claimed gains (76.2 vs 75.0, 67.5 vs 65.7) may reflect faster early convergence of the video-loss-trained parent network rather than a higher final segmentation quality. No OSVOS-VL result is reported at 20k, so the only evidence that the improvement persists at the standard setting is absent. The paper's own note undercuts the central claim without testing it. Per-sequence Table 3 also shows large regressions (car-roundabout 88.3 vs 75.6; scooter-black 58.2 vs 35.4 on VGG16), and results are single-run, so the average gains are not established as robust or statistically significant. The load-bearing assumption is that the advantage survives the 20k setting; this is neither demonstrated nor supported by any reported data.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a 'video loss' module to be added during the parent-network training stage of OSVOS, a fine-tuning-based video object segmentation method. Three variants are introduced: a two-dimensional video loss that uses video identity to separate object instances, a high-dimensional triplet loss, and a mixed instance-aware loss combining contrastive center loss and triplet loss. Experiments on DAVIS-2016 with VGG16 and MobileNet backbones report J-mean improvements after online fine-tuning over the OSVOS baseline (e.g., 76.2 vs. 75.0 for VGG16, 67.5 vs. 65.7 for MobileNet). The paper also claims the module is a common building block generalizable to other fine-tuning methods and to tasks such as depth estimation and saliency detection.","tokens_in":8486,"tokens_out":3963,"duration_ms":37904,"significance":"If the claim holds, the contribution is a simple, lightweight loss that improves instance discrimination during parent-network training, potentially benefiting any fine-tuning-based method. The experimental evidence, however, is weakened by the evaluation protocol: all fine-tuning results are reported at 10k iterations, while the paper itself notes that OSVOS replicates the published numbers only around 20k iterations. No video-loss results are provided at that standard setting, so the claimed gains may reflect faster early convergence rather than a higher final accuracy. Furthermore, the paper does not report several loss hyperparameters, making the method hard to reproduce, and the per-sequence tables show substantial regressions on some videos with single-run results. These issues are addressable with additional experiments and reporting, but they currently prevent the central claim from being considered established.","major_comments":[{"comment":"All fine-tuning results are reported after 10k online iterations, while the paper states in Section 4.2 that around 20k iterations the OSVOS baseline replicates the numbers originally reported in [2]. No OSVOS-VL variants are reported at the 20k setting, so the claimed improvements (76.2 vs. 75.0 for VGG16, 67.5 vs. 65.7 for MobileNet) are established only against an under-trained OSVOS baseline. The central claim of the abstract requires evidence at the standard OSVOS configuration; please provide J-mean at 20k iterations for all proposed variants and baselines.","section":"Section 4.2, Tables 1-2"},{"comment":"The hyperparameters beta1, beta2, and lambda (the margin in the triplet and center losses), as well as the embedding dimension and the number of sampled points per part, are never reported. These are free parameters of the proposed losses, and without them the mixed loss in Eq. (4) is not reproducible. The paper should specify these values and, ideally, provide a sensitivity analysis to show that the reported gains are not tied to a particular tuning.","section":"Sections 3.3-3.4, Eqs. (2)-(4)"},{"comment":"Per-sequence results show large degradations for several sequences (e.g., car-roundabout: 88.3 to 75.6; scooter-black: 58.2 to 35.4 on VGG16). All experiments appear to be single runs, and the paper reports no error bars or multiple seeds. Given these large per-sequence regressions, the average gain of about 1.2 J-mean could easily lie within run-to-run variance. The claim of 'significant improvement' requires either multiple runs or statistical testing.","section":"Section 4.4, Table 3"},{"comment":"The paper repeatedly claims that video loss generalizes to other fine-tuning based methods (e.g., OnVOS, OSVOS-S, CINM) and to similar vision tasks such as depth estimation and saliency detection, but no experiments support these claims. Since this is part of the paper's stated contribution, either provide supporting experiments for at least one other method or task, or explicitly restrict the claims to OSVOS on DAVIS-2016.","section":"Abstract, Section 2.4, and Conclusion"}],"minor_comments":[{"comment":"There are several typos and misspellings, including 'DAVIS' appearing as 'DA VIS' in the abstract and Section 1, 'onlie' for 'online', 'sourrondings' for 'surroundings', and 'dimensioins' for 'dimension'. The paper would benefit from a careful proofreading pass.","section":"Throughout"},{"comment":"The implementation of 2D-VL is unclear: 'we add the identity of each video (vid) into the training process as input' and 'our 2D-VL only updates corresponding (video) category directly' do not specify how the additional head is structured, how the video identity is encoded, or how gradients are routed. Please provide a precise description of the 2D-VL head and its loss computation.","section":"Section 3.2"},{"comment":"The sentence 'For training the parent network, fine-tuning of 240 epochs is conducted based on the initialization of ImageNet features' is ambiguous because the phrase 'fine-tuning' is also used for online fine-tuning later in the same paragraph. Clarify which stage uses 240 epochs and which uses 10k iterations.","section":"Section 3.5"},{"comment":"The caption of Figure 4 does not explain the meaning of the dashed arrows or the 'Pull force' and 'Push force' annotations. Expanding the caption would help the reader understand the proposed high-dimensional video loss.","section":"Figure 4"}],"recommendation":"major_revision","confidential_remarks":"The core idea is plausible and potentially useful, but the evaluation protocol is the central problem: the 10k-iteration setting favors the proposed method if it accelerates early convergence, without demonstrating a higher final performance at the 20k setting used by OSVOS. The authors should be asked to report results at 20k iterations for all variants. If the advantage disappears at 20k, the paper would need to be reframed as a convergence-speed benefit rather than an accuracy improvement. Also, the missing hyperparameter values and single-run results currently make the reproducibility and statistical robustness insufficient for acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this paper's idea is worth a second look, but the experiments don't support its headline. The authors add a video-loss head to OSVOS's parent-network training — a video-id classification branch plus triplet/contrastive losses on embeddings — and show higher J-mean on DAVIS-2016 at 10k online fine-tuning iterations. The problem is the baseline: OSVOS's published numbers are around 20k iterations, and the paper itself notes that ~20k replicates the original OSVOS numbers. At 10k, the OSVOS baseline is under-trained, so the reported gains (76.2 vs 75.0 on VGG16) may be just faster early convergence, not better final quality. No video-loss result at 20k is reported, so the central claim is unestablished.\n\nWhat's genuinely new is the particular combination: using a video-id classification loss alongside pixel-wise metric losses as a plug-in module for parent training. That's a legitimate extension of De Brabandere et al.'s discriminative loss and PML, and it's cheap to try. Testing on two backbones and giving per-sequence numbers is more than many short papers do.\n\nThe soft spots are serious, though. Besides the iteration mismatch, everything is single-run with no error bars; the per-sequence table has some large regressions (car-roundabout, scooter-black on VGG16) that average out but aren't explained; the loss weights beta1, beta2 and lambda are never given; and the claimed generalization to depth and saliency is only mentioned in the abstract with no experiments. The paper also gives no code or trained models. The citation pattern is fine; the related work is covered.\n\nI think the right verdict is a reject, but not a contemptuous one. The core observation — that auxiliary instance-level constraints during parent training can help the later fine-tuning stage — has plausibility and deserves to be tested properly. The authors just need to compare at the standard 20k setting, report multiple runs, and release hyperparameters. If they do, this could become a useful small contribution. As it stands, the evidence doesn't support the abstract's claim of significant improvement.\n\nFor peer review: I wouldn't send this to referees in its current form; the missing comparison is too basic. If the authors fix the protocol, it could be a workshop-level paper or a short at a main venue.","headline":"Reasonable auxiliary-loss idea, but the central comparison is against an under-trained OSVOS baseline, so the headline claim doesn't survive.","tokens_in":9008,"tokens_out":3616,"would_cite":false,"duration_ms":35634,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Adding video loss lifts OSVOS from 75.0 to 76.2 J-mean on DAVIS-2016, and to 67.5 with MobileNet.","keywords":["video object segmentation","OSVOS","video loss","instance-aware learning","online fine-tuning","DAVIS-2016","triplet loss","contrastive center loss"],"falsifier":"Run both methods to 20k online fine-tuning iterations on DAVIS-2016 and compare J-mean; if the video-loss gain vanishes or reverses at that budget, the reported improvement is an artifact of the shortened training protocol.","tokens_in":7994,"feed_emoji":"🎬","tokens_out":6559,"duration_ms":59195,"temperature":0.7,"pith_summary":"This paper argues that the parent-network training phase of one-shot video object segmentation (OSVOS) overemphasizes generic objectness and washes out instance-level cues, which slows the later online fine-tuning step. To fix this, it adds a lightweight video-loss head that supervises the parent network with video identity during training on DAVIS-2016. The paper reports that this module raises the final J-mean on DAVIS-2016 from 75.0 to 76.2 with VGG16 and from 65.7 to 67.5 with MobileNet when the mixed instance-aware video loss is used. If correct, the result matters because video loss is a small, common building block that can be attached to any fine-tuning-based video object segmentation method without changing inference cost.","feed_headline":"Adding video loss lifts OSVOS from 75.0 to 76.2 on DAVIS-2016","feed_subtitle":"Instance-aware video constraints during parent training prepare the network for online fine-tuning on VGG16 and MobileNet.","key_machinery":"The mechanism is the video loss block: a lightweight auxiliary head attached in parallel with the prediction head during parent-network training, supervised by the identity of the training video rather than by the binary foreground/background label alone. It comes in three forms: 2D-VL applies the same weighted cross-entropy expression to per-video categories in image coordinates; HD-VL maps features into a high-dimensional embedding and applies a triplet loss over sampled foreground/background points; and the mixed loss adds a contrastive center loss that penalizes the distance between foreground and background cluster centers. The block works by enforcing that frames of the same video map near one another while different videos map far apart, an 'average object' per sequence, so the parent network keeps instance cues that online fine-tuning can exploit.","core_discovery":"The central claim is that adding video-level constraints to step (2) of OSVOS, training the parent network, produces a parent representation that preserves instance identity, so that step (3), online fine-tuning on the first frame of a test video, adapts faster and more accurately. The paper formalizes the constraint through three losses: a two-dimensional video loss that treats each training video as a category in image space; a high-dimensional triplet loss that pulls foreground pixels together and foreground away from background in an embedding space; and a mixed instance-aware video loss combining triplet loss with a contrastive center loss that keeps the foreground and background cluster centers apart. On DAVIS-2016, OSVOS with the two-dimensional video loss reaches 76.2 J-mean with VGG16 and 66.2 with MobileNet, and the mixed loss reaches 67.5 with MobileNet, all measured after 10k online fine-tuning iterations without post-processing. The authors read this as evidence that instance-level supervision in the parent phase, not any change to the fine-tuning phase, is what removes the noisy, similar-looking false positives that limit OSVOS.","pith_inferences":["The paper compares all methods at 10k online fine-tuning iterations and states that the OSVOS baseline reaches the original reported numbers only around 20k iterations; no video-loss results are reported at 20k, so whether the advantage persists at the canonical budget is untested.","If the mechanism is instance separability in the parent embedding, then the same recipe, discriminative instance-level pretraining before few-shot adaptation, should transfer to other few-shot settings where a base model is fine-tuned to a single exemplar, such as few-shot classification or tracking.","A direct way to test the proposed mechanism would be to measure nearest-neighbor distances in the parent network's embedding: sequences where the target instance is close to a distractor should show the largest improvements, and the per-sequence table already hints at that pattern."],"forward_implications":["On DAVIS-2016, adding the video loss raises the online fine-tuning J-mean from 75.0 to 76.2 with a VGG16 backbone and from 65.7 to 67.5 with MobileNet when the mixed loss is used.","The gains are concentrated in sequences with abrupt motion or distractors that resemble the target object, such as bmx-trees, dog, drift-chicane, and paragliding-launch.","Because video loss is a common module, it should transfer to other fine-tuning-based video object segmentation methods such as OSVOS-S, OnAVOS, and CINM, and to tasks like depth estimation and saliency detection.","The extra training cost is small: the video-loss head runs in parallel with prediction and does not change the online fine-tuning procedure or inference."],"supporting_citations":[{"why":"Defines the OSVOS three-step training baseline that the paper modifies and defends.","marker":"[2]"},{"why":"Provides the DAVIS-2016 benchmark and J-mean evaluation used for all comparisons.","marker":"[17]"},{"why":"Supplies the pixel-wise metric learning triplet formulation adapted for the high-dimensional video loss.","marker":"[5]"},{"why":"Inspires the contrastive center loss used in the mixed instance-aware video loss.","marker":"[8]"},{"why":"Supplies the weighted cross-entropy loss formulation used for both prediction and the 2D video loss.","marker":"[24]"},{"why":"Provides the fully convolutional network architecture that underlies OSVOS and the video-loss experiments.","marker":"[14]"},{"why":"Provides the MobileNet backbone used to show the module generalizes across network structures.","marker":"[11]"},{"why":"One of the fine-tuning-based methods the paper says video loss can generalize to.","marker":"[16]"}],"fun_headline_variants":["Video loss boosts OSVOS segmentation on DAVIS-2016","Instance-aware video loss improves OSVOS on video segmentation","Add video loss to OSVOS parent training for better segmentation","OSVOS gains from video loss in parent phase on DAVIS"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim rests on comparing OSVOS and OSVOS-VL at 10k online fine-tuning iterations; if the improvement disappears when both are run to the 20k iterations needed for OSVOS to reach its originally reported scores, the central claim fails.","fun_headline_variants_meta":{"raw":{"variants":["Video loss boosts OSVOS segmentation on DAVIS-2016","Instance-aware video loss improves OSVOS on video segmentation","Add video loss to OSVOS parent training for better segmentation","OSVOS gains from video loss in parent phase on DAVIS"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000537,"raw_usage":{"total_tokens":2654,"prompt_tokens":1096,"completion_tokens":1558,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":712,"completion_tokens_details":{"reasoning_tokens":1484}},"tokens_in":712,"tokens_out":1558,"duration_ms":10774,"temperature":1.0,"reasoning_tokens":1484,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:37:00.620448+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run both methods to 20k online fine-tuning iterations on DAVIS-2016 and compare J-mean; if the video-loss gain vanishes or reverses at that budget, the reported improvement is an artifact of the shortened training protocol.","supporting_citations":[{"cited_title":"One-shot video object segmentation","cited_arxiv_id":null,"evidence_quote":"Defines the OSVOS three-step training baseline that the paper modifies and defends."},{"cited_title":"A benchmark dataset and evaluation methodology for video object segmentation","cited_arxiv_id":null,"evidence_quote":"Provides the DAVIS-2016 benchmark and J-mean evaluation used for all comparisons."},{"cited_title":"Blazingly fast video object segmentation with pixel-wise metric learning","cited_arxiv_id":null,"evidence_quote":"Supplies the pixel-wise metric learning triplet formulation adapted for the high-dimensional video loss."},{"cited_title":"Holistically-nested edge detection","cited_arxiv_id":null,"evidence_quote":"Supplies the weighted cross-entropy loss formulation used for both prediction and the 2D video loss."},{"cited_title":"Video Object Segmentation Without Temporal Information","cited_arxiv_id":"1709.06031","evidence_quote":"One of the fine-tuning-based methods the paper says video loss can generalize to."}],"review_version":1}