Research map / Video / Video-language understanding
Bias-Conflict Sample Synthesis and Adversarial Removal Debias Strategy for Temporal Sentence Grounding in Video
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes BSSARD, an adversarial debiasing strategy that dynamically synthesizes bias-conflict sample features from visual and query bias generators and trains a span-based grounding model with a bias discriminator to remove dataset bias in temporal sentence grounding.
Problem
Dataset bias in temporal sentence grounding in video (TSGV): uneven temporal distribution of target moments for samples sharing semantic components causes models to rely on spurious shortcuts that hurt out-of-distribution generalization.
Approach
Generate bias-conflict sample features conditioned on random fake target moments using visual and query bias generators, and train the grounding model adversarially with a bias discriminator so it must learn cross-modal alignment rather than spurious single-modality correlations.
Key results
- BSSARD-VSLNet* reaches 47.20 r1i5 and 44.59 mIoU on Charades-CD OOD, versus 43.08 and 41.52 for VSLNet*.
- BSSARD-QAVE improves OOD r1i5 on ActivityNet-CD from 21.39 to 23.11 and on Charades-CD from 37.84 to 44.47.
- BSSARD-EXCL reaches 41.93 r1i5 on Charades-CD OOD versus 39.61 for EXCL.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add BSSARD to VSLNet* | VSLNet* · Charades-CD OOD | r1i5: 43.08 -> 47.20 | Helped |
| Add BSSARD to VSLNet* | VSLNet* · ActivityNet-CD OOD | mIoU: 30.33 -> 31.49 | Helped |
| Add visual and query bias generators (ablation) | VSLNet* · Charades-CD OOD | r1i5: 43.08 -> 47.20 | Helped |
| Use only the query bias generator (ablation) | VSLNet* · Charades-CD OOD | r1i5: 43.08 -> 44.86 | Helped |
| Use only the visual bias generator (ablation) | VSLNet* · Charades-CD OOD | r1i5: 43.08 -> 45.16 | Helped |
| Use visual bias injection before and language bias injection after the feature encoder (ablation) | BSSARD-VSLNet* · Charades-CD OOD | r1i7: 25.30 -> 27.17 | Helped |
| Represent fake temporal position as R^{3 x n} instead of R^n (ablation) | Visual bias generator · Charades-CD OOD | r1i5: 46.40 -> 47.20 | Helped |
| Add the spatial stream F_S to the temporal-position fusion inputs (ablation) | Visual bias generator · Charades-CD OOD | r1i5: 47.20 -> 46.70 | Hurt |
| Alternate the two bias generators each training step rather than each epoch (ablation) | BSSARD-VSLNet* · Charades-CD OOD | r1i5: 46.79 -> 47.20 | Helped |
| Use random selection of a generator each step instead of alternating each step (ablation) | BSSARD-VSLNet* · Charades-CD OOD | r1i5: 47.20 -> 45.81 | Hurt |
| Add KL divergence loss to generation localization loss without discriminator classification loss (ablation) | BSSARD-VSLNet* · Charades-CD OOD | mIoU: 43.49 -> 44.02 | Helped |
| Add KL divergence loss to generation and adversarial losses (ablation) | BSSARD-VSLNet* · Charades-CD OOD | mIoU: 44.00 -> 44.59 | Helped |
Surprising findings
- The visual-only variant VSLNet*-V outperforms the full BSSARD-VSLNet* on the ActivityNet-CD IID test set, attributed to the strong cross-modal alignment of the backbone.
- On Charades-CD IID, BSSARD-VSLNet* falls behind SVTG, which the authors attribute to a higher proportion of biased samples in that IID split.
Limitations
- Gains on IID test sets are mixed; some variants underperform the baseline or SVTG on IID splits.
- The method relies on sampling random fake target moments and only synthesizes features in feature space rather than raw videos.
- Evaluated only on TSGV datasets; generalization to other tasks such as VQA is untested.
Experimental setup
Datasets: Charades-CD ActivityNet-CD
Benchmarks: Charades-CD ActivityNet-CD
Metrics: R@1 IoU=0.5 R@1 IoU=0.7 mIoU
Baselines: 2D-TAN LG DRN VSLNet DCM SVTG MDD QAVE EXCL
Builds on: G^3 VSLNet QAVE EXCL
Closest papers by meaning
- Curriculum Multi-Negative Augmentation for Debiased Video Grounding (AAAI 2023 · similarity 0.87)
- Embracing Uncertainty: Decoupling and De-Bias for Robust Temporal Grounding (CVPR 2021 · similarity 0.86)
- Empower Words: DualGround for Structured Phrase and Sentence-Level Temporal Grounding (NeurIPS 2025 · similarity 0.85)
- Universal Video Temporal Grounding with Generative Multi-modal Large Language Models (NeurIPS 2025 · similarity 0.84)
- SnAG: Scalable and Accurate Video Grounding (CVPR 2024 · similarity 0.84)
- Multi-Pair Temporal Sentence Grounding via Multi-Thread Knowledge Transfer Network (AAAI 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.