Research map / Vision-language and multimodal / Visual question answering
Guessing State Tracking for Visual Dialogue
Wei Pang, Xiaojie Wang
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 early supervision and plain supervision (ES&PS) (ablation) | GST guesser (SL) · GuessWhat?! NewObject, Greedy | guessing success rate (%): 42.58 -> 54.10 | Improved |
| add early supervision and plain supervision (ES&PS) (ablation) | GST guesser (SL) · GuessWhat?! NewGame, Greedy | guessing success rate (%): 39.48 -> 50.97 | Improved |
| add incremental supervision (IS) (ablation) | GST guesser (SL) · GuessWhat?! NewObject, Greedy | guessing success rate (%): 53.49 -> 54.10 | Improved |
| add incremental supervision (IS) (ablation) | GST guesser (SL) · GuessWhat?! NewGame, Greedy | guessing success rate (%): 50.33 -> 50.97 | Improved |
| use symmetric concatenation [h_qa; h_qa*O; O] instead of [h_qa; O] (ablation) | GST guesser (SL) · GuessWhat?! test | error rate (%): 37.8 -> 34.4 | Improved |
| use symmetric concatenation instead of only element-wise product (ablation) | GST guesser (SL) · GuessWhat?! test | error rate (%): 36.7 -> 34.4 | Improved |
| raise c in the log term from 1.1 to 1.5 or 2.0 (ablation) | GST guesser (SL) · GuessWhat?! test | error rate (%): 34.3 -> 34.8 | No clear effect |
| replace single-step guesser with guessing state tracking (GST), VDST QGen, 8 questions, RL | guesser (previous) · GuessWhat?! NewObject, Greedy | guessing success rate (%): 70.55 -> 83.32 | Improved |
| replace single-step guesser with GST, VDST QGen, 5 questions, SL | guesser (previous) · GuessWhat?! NewObject, Greedy | guessing success rate (%): 49.49 -> 53.35 | Improved |
| replace single-step guesser with GST, qgen, 5 questions, RL | guesser (previous) · GuessWhat?! NewObject, Greedy | guessing success rate (%): 60.3 -> 67.06 | Improved |
| increase maximum questions from 5 to 8 for GST, RL | GST + VDST · GuessWhat?! NewObject, Greedy | guessing success rate (%): 77.30 -> 83.32 | Improved |
| GST trained by RL instead of SL (8 questions) | GST · GuessWhat?! test | error rate (%): 34.3 -> 18.4 | Improved |
Closest papers by meaning
- Visual Dialogue State Tracking for Question Generation (AAAI 2020 · similarity 0.94)
- Learning Better Visual Dialog Agents With Pretrained Visual-Linguistic Representation (CVPR 2021 · similarity 0.91)
- GuessWhat?! Visual Object Discovery Through Multi-Modal Dialogue (CVPR 2017 · similarity 0.89)
- Divide-and-Conquer: Tree-structured Strategy with Answer Distribution Estimator for Goal-Oriented Visual Dialogue (AAAI 2025 · similarity 0.86)
- DialogueVPR: Towards Conversational Visual Place Recognition (CVPR 2026 · similarity 0.84)
- RefDetector: A Simple Yet Effective Matching-based Method for Referring Expression Comprehension (AAAI 2025 · similarity 0.84)
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.
