Research map / Video / Video-language understanding
TimeCraft: Navigate Weakly-Supervised Temporal Grounded Video Question Answering via Bi-directional Reasoning
Huabin Liu, Xiao Ma, Cheng Zhong, Yang Zhang, Weiyao Lin
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the full bi-directional reasoning framework (dual Q&A, reconstruction and grounding consistency) (ablation) | Temp[CLIP] · NExT-GQA test | Acc@GQA: 15.5 -> 18.2 | Improved |
| add dual Q&A data to standard QA training (ablation) | Temp[CLIP] · NExT-GQA test | Acc@QA: 59.4 -> 60.8 | Improved |
| add reconstruction loss L_recon (ablation) | TimeCraft · NExT-GQA test | mIoU: 9.2 -> 12.1 | Improved |
| add grounding consistency loss L_ground without reconstruction (ablation) | TimeCraft · NExT-GQA test | Acc@GQA: 15.8 -> 15.6 | No clear effect |
| add both L_recon and L_ground on top of dual Q&A (ablation) | TimeCraft · NExT-GQA test | mIoU: 12.1 -> 15.6 | Improved |
| use Llama-2 13B instead of Gemini or InternLM for dual Q&A (ablation) | dual Q&A generation · NExT-GQA test | Acc@GQA: 17.8 -> 18.2 (Gemini) | Improved |
| add dual QA as augmentation data to NG+ (ablation) | NG+ (Temp[CLIP]) · NExT-GQA test | Acc@GQA: 16.0 -> 16.3 | Improved |
| use question-critical grounding instead of answer-critical grounding (ablation) | TimeCraft · NExT-GQA test (causal) | mIoU: 17.4 -> 15.9 | Hurt |
| use union of answer-critical and question-critical grounding (ablation) | TimeCraft · NExT-GQA test (causal) | mIoU: 17.4 -> 18.6 (causal mIoU); 13.8 -> 12.9 (temporal mIoU) | Mixed |
| increase sampled frames to 128 (ablation) | TimeCraft · NExT-GQA test | mIoP and mIoU | Hurt |
| apply TimeCraft training to FrozenBiLM vs Temp[CLIP] | TimeCraft · NExT-GQA test | Acc@QA: 73.1 -> 74.7 (FrozenBiLM); 63.3 -> 65.6 (Temp[CLIP]) | Mixed |
| apply bi-directional training to Env-QA | Temp[CLIP] · Env-QA test | accuracy (All): 48.97 -> 49.02 (vs MIST[CLIP]) | No clear effect |
Later papers that compared with this method
1 later accepted paper reports TimeCraft as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Wavelet-based Intra-video Counterfactual Reasoning for Video Question Grounding (ECCV 2026 · better on 3 of 3 shared results)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Commonsense Video Question Answering through Video-Grounded Entailment Tree Reasoning (CVPR 2025 · similarity 0.92)
- MoReVQA: Exploring Modular Reasoning Models for Video Question Answering (CVPR 2024 · similarity 0.91)
- Can I Trust Your Answer? Visually Grounded Video Question Answering (CVPR 2024 · similarity 0.91)
- Searching Videos as Trees: Self-Correcting Agents for Grounded Long Video QA (NeurIPS 2026 · similarity 0.90)
- Grounded Multi-Hop VideoQA in Long-Form Egocentric Videos (AAAI 2025 · similarity 0.90)
- Modularized Self-Reflected Video Reasoner for Multimodal LLM with Application to Video Question Answering (ICML 2025 · similarity 0.90)
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.
