Research map / Label-efficient and robust learning / Domain adaptation and generalization
Beyond Sharing Weights for Deep Domain Adaptation
Artem Rozantsev, Mathieu Salzmann, Pascal Fua
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 MMD loss to the two-stream loss (ablation) | Ours (Ls+Lt+Lw) · UAV-200 (full, supervised DA) | AP: 0.673 -> 0.732 | Improved |
| add weight-linking term Lw to Ls+Lt+LMMD (ablation) | Ours · UAV-200 (full, supervised DA) | AP: 0.711 -> 0.732 | Improved |
| use non-shared weights with regularizer instead of shared weights (DDC) | two-stream CNN · UAV-200 (full, supervised DA) | AP: 0.664 -> 0.732 | Improved |
| add synthetic data pre-training and fine-tuning on real data (vs real only) | CNN · UAV-200 (full, supervised DA) | AP: 0.575 -> 0.732 | Improved |
| naively train on real and synthetic together (Ls+Lt) | CNN · UAV-200 (full, supervised DA) | AP: 0.575 -> 0.569 | No clear effect |
| replace non-shared weights by DDC shared weights | two-stream CNN · UAV-200 (small, supervised DA) | accuracy: 0.92 -> 0.89 | Hurt |
| drop weight-linking terms while still minimizing MMD (ablation) | Ours · UAV-200 (full, supervised DA) | AP: 0.732 -> 0.711 | Hurt |
| choose layer sharing by lowest MMD instead of best validation architecture | UAV architecture selection · UAV-200 (small, supervised DA) | test accuracy: 0.757 -> 0.732 | No clear effect |
| allow non-shared weights in last two layers vs shared (GRL) | two-stream AlexNet · Office A->W | accuracy: 0.730 -> 0.760 | Improved |
| use non-shared weights instead of shared (DDC) | two-stream CNN · MNIST-USPS (average) | accuracy: 0.554 -> 0.640 | Improved |
| use non-shared weights instead of shared (DDC) | two-stream regressor · Facial landmark pose estimation | average PCP score (%): 60.3 -> 62.8 | Improved |
| add more real positive samples (200 to 5000) with 32800 synthetic | Ours · UAV-200 (full, supervised DA) | AP | Improved |
Closest papers by meaning
- Unsupervised Domain Adaptation With Similarity Learning (CVPR 2018 · similarity 0.91)
- Unsupervised Domain Adaptation of Object Detectors: A Survey (TPAMI 2023 · similarity 0.90)
- Boosting Domain Adaptation by Discovering Latent Domains (CVPR 2018 · similarity 0.90)
- MultiDIAL: Domain Alignment Layers for (Multisource) Unsupervised Domain Adaptation (TPAMI 2020 · similarity 0.90)
- Unsupervised Domain Adaptation with Residual Transfer Networks (NeurIPS 2016 · similarity 0.90)
- Unsupervised Domain Adaptation Using Feature-Whitening and Consensus Loss (CVPR 2019 · similarity 0.90)
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.
