REVIEW 2 major objections 2 minor 1 cited by
From Waveforms to Pixels: A Survey on Audio-Visual Segmentation
T0 review · 2 major / 2 minor · reviewed 2026-08-06 · deepseek-v4-flash
Pith's one-line read Locating and outlining every object that is making the sound in a video has become a research field large enough for a single systematic survey, and this paper is that survey.
desk verdict As submitted, this is an abstract with no body text, so the survey's central claims cannot be checked and the paper is not ready for peer review. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is the audio-visual segmentation task itself, defined as the pixel-level delineation of every object in a video that is producing the sound in the accompanying audio track. The survey's machinery is a three-axis taxonomy of methods: the choice of unimodal versus multimodal encoding, the strategy used to fuse audio and visual features, and the design of the decoder that turns fused features into segmentation masks. A fourth axis crosses these with training paradigms, ranging from fully supervised through weakly supervised to training-free setups. Benchmark datasets and evaluation metrics do the measuring work: they are what allow the survey to compare methods and to attribute performance differences to architectural choices, fusion strategies, and training regimes.
What would settle it
Collect the reported results for one standard benchmark and check each method's evaluation protocol—exact split, metric definition, post-processing, and whether audio is used at test time. If the numbers mix incompatible protocols, or if a spot re-run of the top methods changes their ordering, the survey's comparative conclusions fail.
Extended reading notes
Core claim
The central claim of this survey is that audio-visual segmentation has become a distinct, mature research area whose entire arc—from problem formulation through datasets, metrics, architectures, fusion strategies, decoder designs, and training paradigms—can be laid out in a single coherent narrative. On the paper's own terms, the discovery is taxonomic: methods are not a scattered set of tricks but a progression along identifiable axes, and their relative standing can be judged through extensive comparison across standard benchmarks. The survey further asserts that the field's current limits are concentrated in specific, named weaknesses, and that the remedies are already visible in weakly supervised and training-free approaches, foundation-model generalization, and higher-level reasoning. This is a reference-map claim rather than a new algorithm, and its force lies in how completely and fairly it organizes what has been done.
Load-bearing premise
The survey's value rests on its literature selection and benchmark numbers being accurate, representative, and compared under consistent conditions; if key methods are missing or scores come from different evaluation protocols, the field map and its rankings mislead.
Editorial extensions
If this is right
- A newcomer to AVS can use the problem formulation, dataset descriptions, and metric discussion as a direct entry point for designing a first method or benchmark study.
- Because the survey compares methods under standard benchmarks, future work can pick the strongest encoding, fusion, and decoder configurations as baselines rather than re-deriving them.
- The survey's challenge list—temporal modeling, vision bias, robustness, computation—effectively defines a near-term research agenda for the field.
- The progression from fully supervised to weakly supervised and training-free methods indicates that reducing reliance on labeled data is a live and promising direction.
- If the identified limitations are accurate, approaches that strengthen temporal reasoning and audio-visual fusion in complex scenes should expect the largest performance gains.
Reading between the lines
- A testable extension follows from the reported vision bias: design audio-first or audio-only probes that attempt segmentation without visual features; if they recover significant structure, current methods may be leaning on vision more than the fusion narrative suggests.
- The survey's emphasis on fusion implies that a deliberately simple, parameter-free fusion baseline—such as broadcasting audio features to every pixel and concatenating—would be a cheap way to expose how much of the reported gains come from architecture sophistication rather than training or data.
- The future direction of higher-level reasoning suggests that AVS will likely converge with audio-visual language grounding; one concrete prediction is that models trained with language or event-level supervision will outperform pixel-level fusion methods on rare or novel sound sources.
- Because the survey covers training-free methods, an immediate benchmark extension would be to test foundation-model segmenters prompted by audio-derived text descriptions; the survey's comparison framework could be reused to score such zero-shot pipelines.
Signed reviews
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The manuscript, as submitted, consists solely of an abstract and an empty full-text body. The abstract claims to present a comprehensive survey of audio-visual segmentation (AVS), covering problem formulation, benchmark datasets, evaluation metrics, methodological progression, training paradigms, and benchmark comparisons, and to close with challenges and future directions. Because no body text, figures, tables, or references are provided, the claims in the abstract cannot be checked or assessed.
Significance. If the full survey were present and the content matched the abstract's claims, the paper would serve as a potentially valuable reference map for the growing AVS community, consolidating method taxonomies, benchmarking protocols, and open problems. The abstract indicates a clear scope and an organized structure, and the promised extensive comparison across standard benchmarks would be a practical contribution. However, with the full text absent, the current submission has no verifiable content, and thus its significance is entirely contingent on the completion and accuracy of the missing material.
major comments (2)
- [Full Text (entire body)] The submitted manuscript contains no body text whatsoever; it is only an abstract. The central claim of the paper—that it provides a 'comprehensive overview' and an 'extensive comparison' of AVS methods—cannot be verified, replicated, or even located in the submission. There are no sections describing problem formulation, datasets, evaluation metrics, methodology taxonomy, benchmark tables, or discussions of training paradigms, despite each being promised in the abstract. This is a load-bearing omission: a survey paper's value resides entirely in its detailed exposition and evidence, none of which is present. The authors must provide the full manuscript before any substantive review can proceed.
- [Abstract, 'extensive comparison'] The abstract asserts that the paper includes 'an extensive comparison of AVS methods across standard benchmarks, highlighting the impact of different architectural choices, fusion strategies, and training paradigms.' In the absence of the corresponding tables, experimental protocols, and cited sources, this assertion is unsupported. No benchmark names, metric definitions, or numerical results are given in the available material, so the reader cannot judge whether the comparison is fair, consistent, or representative.
minor comments (2)
- [Abstract, line on learning paradigms] The phrase 'selfand weakly supervised learning' contains a typographical error; it should be 'self-supervised and weakly supervised learning' or the like.
- [Title and abstract] The title promises a journey 'From Waveforms to Pixels,' but the abstract does not discuss waveform-level audio representations or how they are processed prior to visual alignment. If the full text covers audio feature extraction from waveforms, the abstract should mention it for coherence.
Circularity Check
No circularity identified: the available manuscript is an abstract-only survey with no derivation chain, fitted parameters, or predictions to reduce.
full rationale
The submitted material contains only the abstract and no body text, equations, benchmark tables, or method descriptions. A circularity analysis requires a concrete derivation chain or a fitted input that is later renamed as a prediction; here there is no such chain present. The abstract makes survey-level claims of comprehensiveness and extensive benchmarking, but those claims are unverifiable from the available text rather than circular: nothing in the abstract defines a quantity in terms of itself, fits a parameter and calls it a prediction, or imports a load-bearing result solely from the authors' prior work. The absence of full text is a verification gap, not a circularity finding. Accordingly, the honest verdict is no significant circularity, with score 0.
Assumptions & free parameters
assumptions (1)
- domain assumption The authors' selection and characterization of prior AVS work is accurate and representative.
Cite this review
Pith. "Pith review of From Waveforms to Pixels: A Survey on Audio-Visual Segmentation." pith.science (2026). https://pith.science/paper/64H5YCYK
@misc{pith2026250803724,
author = {Pith},
title = {Pith review of: From Waveforms to Pixels: A Survey on Audio-Visual Segmentation},
year = {2026},
howpublished = {\url{https://pith.science/paper/64H5YCYK}},
note = {Machine review of arXiv:2508.03724}
}
read the original abstract
Audio-Visual Segmentation (AVS) aims to identify and segment sound-producing objects in videos by leveraging both visual and audio modalities. It has emerged as a significant research area in multimodal perception, enabling fine-grained object-level understanding. In this survey, we present a comprehensive overview of the AVS field, covering its problem formulation, benchmark datasets, evaluation metrics, and the progression of methodologies. We analyze a wide range of approaches, including architectures for unimodal and multimodal encoding, key strategies for audio-visual fusion, and various decoder designs. Furthermore, we examine major training paradigms, from fully supervised learning to weakly supervised and training-free methods. Notably, we provide an extensive comparison of AVS methods across standard benchmarks, highlighting the impact of different architectural choices, fusion strategies, and training paradigms on performance. Finally, we outline the current challenges, such as limited temporal modeling, modality bias toward vision, lack of robustness in complex environments, and high computational demands, and propose promising future directions, including improving temporal reasoning and multimodal fusion, leveraging foundation models for better generalization and few-shot learning, reducing reliance on labeled data through selfand weakly supervised learning, and incorporating higher-level reasoning for more intelligent AVS systems.
Forward citations
Cited by 1 Pith paper
-
Delayed Bidirectional Alignment via Disentangled Audio Semantics for Audio-Visual Segmentation
A delayed bidirectional audio-visual alignment framework with bank-grounded disentangled audio queries achieves state-of-the-art segmentation results on AVS-Objects and VPO benchmarks.
Reviewed August 6, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.