Research map / Vision-language and multimodal / Cross-modal learning
Cross-Modal Knowledge Transfer without Task-Relevant Source Data
Sk Miraj Ahmed, Suhas Lohit, Kuan-Chuan Peng, Michael J. Jones, Amit K. Roy-Chowdhury
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 modality-specific losses (L_d and L_TI) to SOCKET over SHOT | SHOT · SUN RGB-D single source, Kinect v1 target | accuracy (%): 11.7 -> 18.4 | Improved |
| add modality-specific losses to DECISION for two sources | DECISION · SUN RGB-D two sources, Kinect v2 target | accuracy (%): 26.7 -> 31.4 | Improved |
| use SOCKET instead of SHOT on RGB to NIR | SHOT · RGB-NIR Scene RGB to NIR | accuracy (%): 86.7 -> 90.2 | Improved |
| use SOCKET instead of SHOT on NIR to RGB | SHOT · RGB-NIR Scene NIR to RGB | accuracy (%): 92.2 -> 92.7 | Improved |
| use TI data from a different dataset (SUN RGB-D) rather than DIML | SOCKET · DIML RGB+D RGB to depth | accuracy (%): 46.1 -> 53.2 | Improved |
| add discrepancy loss L_d to L_ma (ablation) | SOCKET · SUN RGB-D (a) single source RGB to Kinect v2 depth | accuracy (%): 30.0 -> 31.6 | Improved |
| add TI-data loss L_TI to L_ma (ablation) | SOCKET · SUN RGB-D (a) single source RGB to Kinect v2 depth | accuracy (%): 30.0 -> 34.9 | Improved |
| add both L_d and L_TI to L_ma (ablation) | SOCKET · SUN RGB-D (a) single source RGB to Kinect v2 depth | accuracy (%): 30.0 -> 36.3 | Improved |
| add both L_d and L_TI to L_ma (ablation) | SOCKET · SUN RGB-D (b) two sources kv1+xtion to Kinect v1 depth | accuracy (%): 11.7 -> 23.9 | Improved |
| increase TI images per class from 0 to 60 (ablation) | SOCKET · SUN RGB-D Kinect v1 RGB to depth, six TI classes | accuracy (%): 16.7 -> 25.0 | Improved |
| raise regularization lambda from 0 to 0.5 (ablation) | SOCKET · Kinect v1 RGB to Kinect v1 depth | accuracy (%): 16.1 -> 23.4 | Improved |
| raise regularization lambda from 0.5 to 1.0 (ablation) | SOCKET · Kinect v2 RGB to Kinect v1 depth | accuracy (%): 36.7 -> 16.3 | Hurt |
Closest papers by meaning
- Decoupled and Reusable Adaptation for Efficient Cross-Modal Transfer (CVPR 2026 · similarity 0.88)
- Knowledge As Priors: Cross-Modal Knowledge Generalization for Datasets Without Superior Knowledge (CVPR 2020 · similarity 0.87)
- Asymmetric Cross-Modal Knowledge Distillation: Bridging Modalities with Weak Semantic Consistency (AAAI 2026 · similarity 0.87)
- Learning Modality Knowledge Alignment for Cross-Modality Transfer (ICML 2024 · similarity 0.86)
- C2KD: Bridging the Modality Gap for Cross-Modal Knowledge Distillation (CVPR 2024 · similarity 0.84)
- Knowledge Transfer from Vision Foundation Models for Efficient Training of Small Task-specific Models (ICML 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.
