Research map / Label-efficient and robust learning / Domain adaptation and generalization
Domain Adaptation Through Task Distillation
Brady Zhou, Nimit Kalra, Philipp Krähenbühl
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 |
|---|---|---|---|
| use map-view instead of camera-view proxy | task distillation · SuperTuxKart to CARLA | avg distance traveled (m): 164.6 -> 260.5 | Improved |
| use map-view instead of camera-view proxy | modular pipeline · SuperTuxKart to CARLA | avg distance traveled (m): 110.4 -> 135.3 | No clear effect |
| use task distillation instead of modular transfer with map-view | modular pipeline · SuperTuxKart to CARLA | avg distance traveled (m): 135.3 -> 260.5 | Improved |
| use task distillation instead of modular transfer with map-view | modular pipeline · ViZDoom to CARLA | avg distance traveled (m): 145.3 -> 277.3 | Improved |
| use task distillation instead of modular transfer with camera-view | modular pipeline · ViZDoom to CARLA | avg distance traveled (m): 140.4 -> 166.9 | Improved |
| train driving policy on ground-truth proxy labels instead of predicted proxy labels (ablation) | modular pipeline, map-view · SuperTuxKart to CARLA | avg distance traveled (m): 49.9 -> 135.3 | Improved |
| train driving policy on ground-truth proxy labels instead of predicted proxy labels (ablation) | modular pipeline, cam-view · ViZDoom to CARLA | avg distance traveled (m): 148.6 -> 140.4 | Hurt |
| transfer waypoints instead of raw steering and throttle (ablation) | CyCADA · SuperTuxKart to CARLA | avg distance traveled (m): 26.7 -> 24.0 | No clear effect |
| add task distillation via depth proxy to direct transfer | DeepLabv3 · SYNTHIA-SF to CARLA | mIoU: 21.2 -> 55.2 | Improved |
| add task distillation via depth proxy to direct transfer | DeepLabv3 · SYNTHIA-SF to Cityscapes | mIoU: 15.0 -> 28.0 | Improved |
| use task distillation with DeepLabv3 instead of AT/DT | AT/DT · SYNTHIA-SF to CARLA | mIoU: 43.1 -> 55.2 | Improved |
| use task distillation instead of AT/DT | AT/DT · SYNTHIA-SF to Cityscapes | mIoU: 23.2 -> 28.0 | Improved |
Closest papers by meaning
- Learning Across Tasks and Domains (ICCV 2019 · similarity 0.88)
- Towards Optimal Strategies for Training Self-Driving Perception Models in Simulation (NeurIPS 2021 · similarity 0.88)
- Transferable Representation Learning with Deep Adaptation Networks (TPAMI 2018 · similarity 0.88)
- Towards Universal Dataset Distillation via Task-Driven Diffusion (CVPR 2025 · similarity 0.86)
- Dataset Distillation for Pre-Trained Self-Supervised Vision Models (NeurIPS 2025 · similarity 0.85)
- Knowledge Distillation from Single-Task Teachers to Multi-Task Student for End-to-End Autonomous Driving (AAAI 2024 · similarity 0.85)
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.
