Research map / Vision-language and multimodal / Visual question answering
Mining Interpretable AOG Representations From Convolutional Networks via Active Question Answering
Quanshi Zhang, Jie Ren, Ge Huang, Ruiming Cao, Ying Nian Wu, Song Chun Zhu
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 active question-answering to AOG learning (vs AOG without QA) (ablation) | AOG · CUB200-2011 (head part) | normalized distance: 0.1084 -> 0.0434 | Improved |
| add active QA with only 10 part annotations vs AOG w/o QA with 20 (ablation) | AOG · CUB200-2011 (head part) | normalized distance: 0.1084 -> 0.0626 | Improved |
| increase part annotations from 10 to 20 (ablation) | Ours · CUB200-2011 (head part) | normalized distance: 0.0626 -> 0.0434 | Improved |
| increase part annotations from 10 to 30 (ablation) | Ours · ILSVRC 2013 DET Animal-Part | normalized distance: 0.1126 -> 0.0909 | Improved |
| increase part annotations from 20 to 30 (ablation) | Ours · ILSVRC 2013 DET Animal-Part | normalized distance: 0.0965 -> 0.0909 | Improved |
| increase part annotations from 10 to 30 for Ours+fastRCNN (ablation) | Ours+fastRCNN · PASCAL VOC Part (PCP) | PCP (%): 33.0 -> 50.5 | Improved |
| add second-stage fine-tuning with object boxes to Fast-RCNN (ablation) | Fast-RCNN · CUB200-2011 (head part) | normalized distance: 0.3105 -> 0.1989 | Improved |
| add object-box fine-tuning to CNN-PDD (ablation) | CNN-PDD · CUB200-2011 (head part) | normalized distance: 0.2446 -> 0.2694 | Hurt |
| use RBF-SVM instead of linear SVM on fc7 features (ablation) | fc7 features · CUB200-2011 (head part) | normalized distance: 0.2786 -> 0.3360 | Hurt |
| add fast-RCNN refinement of part localization with 10 annotations | Ours · PASCAL VOC Part (PCP) | PCP (%): 33.0 vs 34.5 | No clear effect |
| more part annotations grow AOG semantic-part children but not part templates (ablation) | AOG · Pascal VOC Part | average number of AOG node children (layer 1): 3.15 -> 15.83 | Mixed |
| use Ours (10 annotations) instead of Fast-RCNN (2 fts) with 10 annotations | Fast-RCNN (2 fts) · Pascal VOC Part (neck) | normalized distance: 0.309 -> 0.152 | Improved |
Closest papers by meaning
- Mining Object Parts From CNNs via Active Question-Answering (CVPR 2017 · similarity 0.96)
- Growing Interpretable Part Graphs on ConvNets via Multi-Shot Learning (AAAI 2017 · similarity 0.88)
- Extraction of an Explanatory Graph to Interpret a CNN (TPAMI 2020 · similarity 0.86)
- Interpreting CNN Knowledge via an Explanatory Graph (AAAI 2018 · similarity 0.85)
- Exploring the Effectiveness of Object-Centric Representations in Visual Question Answering: Comparative Insights with Foundation Models (ICLR 2025 · similarity 0.85)
- Detection-Based Intermediate Supervision for Visual Question Answering (AAAI 2024 · 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.
