Research map / Vision-language and multimodal / Cross-modal learning
Learning with Privileged Information via Adversarial Discriminative Modality Distillation
Nuno C. Garcia, Pietro Morerio, Vittorio Murino
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 128-d bottleneck to the discriminator input (ablation) | hallucination network · NTU full (cross-subject) | accuracy: 54.25% -> 72.14% | Improved |
| use pool+conv bottleneck instead of one conv (ablation) | hallucination network · NTU-mini (cross-subject) | accuracy: 55.64% -> 61.41% | Improved |
| use pool+conv bottleneck instead of spatial conv + 1D conv (ablation) | hallucination network · NTU-mini (cross-subject) | accuracy: 53.21% -> 61.41% | Improved |
| add class (y-hat) classification task to discriminator instead of binary real/fake (ablation) | hallucination network · NTU-mini (cross-subject) | accuracy: 1.81% -> 59.87% | Improved |
| add frame index yt to discriminator input (ablation) | hallucination network · NTU-mini (cross-subject) | accuracy: 59.87% -> 63.03% | Improved |
| add adversarial discriminator with classification task instead of hallucination model of [11] (ablation) | hallucination network · NYUD object recognition | accuracy: 55.94% -> 57.52% | Improved |
| use naive adversarial discriminator (binary task only) (ablation) | hallucination network · NYUD object recognition | accuracy: 57.52% -> 50.81% | Hurt |
| hallucinate depth features instead of estimating depth maps with FCRN (ablation) | RGB-only model · NYUD object recognition | accuracy: 50.23% -> 57.52% | Improved |
| fuse two RGB networks as ensemble (ablation) | RGB stream · NYUD object recognition | accuracy: 52.90% -> 54.14% | Improved |
| fine-tune the fused two-stream model (ablation) | two-stream · NYUD object recognition | accuracy: 57.39% -> 58.73% | Improved |
| add Gaussian speckle noise (variance 10^-1) to depth at test time (ablation) | two-stream · NTU RGB+D (cross-view) | accuracy: 85.49% -> 68.99% | Hurt |
| add hallucination stream to RGB stream at test time (ablation) | ADMD · NTU RGB+D (cross-view) | accuracy: 80.01% -> 81.50% | Improved |
Closest papers by meaning
- Modality Distillation with Multiple Stream Networks for Action Recognition (ECCV 2018 · similarity 0.90)
- Learning With Side Information Through Modality Hallucination (CVPR 2016 · similarity 0.87)
- Deep Multimodal Feature Analysis for Action Recognition in RGB+D Videos (TPAMI 2017 · similarity 0.86)
- Cooperative Training of Deep Aggregation Networks for RGB-D Action Recognition (AAAI 2018 · similarity 0.86)
- Graph Distillation for Action Detection with Privileged Modalities (ECCV 2018 · similarity 0.86)
- Visual Recognition in RGB Images and Videos by Learning from RGB-D Data (TPAMI 2017 · similarity 0.86)
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.
