- Rating
- 1
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 1
- Soundness
- 1
Read review
Paper summary
SQRAct proposes a temporal action detector combining semantic-conditioned frequency filtering, multiscale feature enhancement, and inter-layer query refinement. Building on DiGIT’s decoder, it uses containment and adjacency relationships between predicted segments to refine their representations and boundaries, targeting duplicate and fragmented detections.
Critical strengths
The paper addresses a meaningful distinction between genuine action boundaries and feature changes caused by internal motion or background variation. Explicitly examining relationships between partial predictions is a reasonable direction for improving temporal localization.
Critical weaknesses
-
After reading your manuscript, I think that the claimed foreground/background semantics are unsubstantiated. As shown in Section 3.3 that introduces an action probability without specifying its supervision, loss, or evidence that it identifies foreground. If is constant, both weighted averages in Eq. (1) reduce to the same global feature mean before their separate projections; no stated constraint excludes this degenerate case. Consequently, the central claim of semantic foreground/background conditioning rests on an unverified interpretation of learned weights.
-
Also, in the methodology section, the frequency mechanism does not establish selective suppression of background changes. A single video-conditioned applies the same gain to action and background contributions occupying the same frequency component, so pooled semantic conditioning alone does not resolve their ambiguity. Moreover, Eq. (2) gives , while approaches 2 when its Gaussian term vanishes: these branches amplify rather than individually suppress out-of-band content. Neither the unspecified filter combination nor the experiments demonstrate how this construction preserves genuine boundaries while rejecting visually similar nuisance transitions.
-
QRR substitutes semantic similarity for instance identity without validating that assumption. Consecutive repetitions of the same action can be highly similar yet belong to distinct instances, while different phases of one action can have dissimilar features; the similarity gate therefore does not reliably distinguish fragmentation from legitimate instance separation. Equation (10) permits only expansion within the adjacency update, creating a concrete risk of bridging distinct instances. Containment refinement also moves predictions toward a query selected by classification confidence, although classification confidence does not establish localization quality, a distinction explicitly addressed by ReAct [2], and no evidence shows that this reduces duplicates rather than making them more alike.
-
Another question is that section 3.4 does not formally define the containment, proximity, and adjacency functions, and the paper omits the complete training objective, essential hyperparameters, and a reproducible specification of SFE. Equations (8)–(9) also associate with , despite the prose identifying as the predecessor, leaving the directional convention inconsistent. The appendix promised for essential parameterization is absent from the supplied PDF, so these omissions directly obstruct verification. The central algorithm cannot be reproduced reliably.
Some issues regarding your experimental part.
The main experimental claims are internally inconsistent and use an unexplained weak baseline. Section 4.1 claims 41.5% ActivityNet mAP and a 0.3-point improvement, whereas Table 1 labels the proposed method “FRQAct” and reports 39.2% versus 37.8%, implying a 1.4-point improvement. The original DiGIT reports 42.0% with InternVideo2, or 41.3% without NMS [1], making the unexplained 37.8% reproduction a serious comparison problem. Together with the missing HACS results, these discrepancies leave the claimed superiority across datasets unsupported.
Furthermore, no experiment establishes which proposed mechanism actually contributes. The supplied paper contains no component ablations, repeated-run variability, or direct measurements of boundary errors, duplication, and fragmentation. Given the inherited DiGIT decoder [1], existing relational action queries [2], and prior semantic-adaptive frequency filtering [3], aggregate mAP cannot establish the specific contribution of the proposed combination. A 0.9-point THUMOS14 gain without controlled ablations or capacity and runtime comparisons is inadequate evidence for the necessity of this substantially expanded architecture.
[1] Kim, H.-J., Lee, Y., Hong, J.-H. and Lee, S.-W., 2025, June. DiGIT: Multi-dilated gated encoder and central-adjacent region integrated decoder for temporal action detection transformer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) (pp. 24286–24296).
[2] Shi, D., Zhong, Y., Cao, Q., Zhang, J., Ma, L., Li, J. and Tao, D., 2022, October. ReAct: Temporal action detection with relational queries. ECCV 2022 (pp. 105–121). Springer.
[3] Huang, Z., Zhang, Z., Lan, C., Zha, Z.-J., Lu, Y. and Guo, B., 2023, October. Adaptive frequency filters as efficient global token mixers. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) (pp. 6049–6059).
Questions for authors
-
What supervises , and what evidence excludes constant or semantically incorrect foreground/background assignments? Can GFSF distinguish true boundaries from background changes with overlapping frequency content? Also, why does approach 2 outside its target band, and how are the filter responses combined and constrained?
-
How does QRR distinguish adjacent instances of the same class from fragments of one instance? and why should classification confidence identify the query with better boundaries, and does containment refinement actually reduce duplicate detections?
-
What are the exact relation functions, thresholds, self-pair masks, and rules for queries without eligible neighbors? Also, which predecessor/successor indexing in Eqs. (8)–(10) is correct, and how are conflicting reverse updates resolved?
-
What are the complete training losses, SFE operations, FFT conventions, and training/inference settings?
-
Which ActivityNet results are correct, and why does the reproduced DiGIT fall substantially below its original results? Are feature extraction, external classification scores, and NMS settings identical across the relevant comparisons? What do individual and combined module ablations show, including variation across seeds and computational overhead?