Parameter-Efficient Adaptation of Self-supervised Vision Backbones Survives Distribution Shift in Radar Micro-Doppler HAR
Organizations: ETIS UMR 8051, CY Cergy Paris Université, ENSEA, CNRS, Cergy, France · Fondation Léonie Chaptal, Sarcelles, France · ETIS UMR 8051, CY Cergy Paris Universite, ENSEA, CNRS, F-95000 Cergy, France · Fondation Léonie Chaptal, F-95200 Sarcelles, France
Abstract
Micro-Doppler radar enables contactless, privacy-preserving human activity recognition (HAR), but deployment robustness remains underexplored. We compared full fine-tuning, frozen probing, and parameter-efficient fine-tuning (PEFT) using subject-disjoint in-distribution cross-validation and few-shot cross-age, cross-orientation, and cross-dataset transfer. Full fine-tuning suffered negative cross-dataset transfer, while frozen features under-adapted on cross-orientation. Only PEFT avoided both failures. We paired DINOv3 with all-layer principal-singular-direction adaptation (PiSSA), to our knowledge the first application of a self-supervised image foundation backbone to radar micro-Doppler HAR. DINOv3-PiSSA ranked first among PEFT methods in every evaluated setting. Under our protocol, its Small variant matched the in-distribution accuracy of the prior Base-size PEFT state of the art using 4.5x fewer total parameters, with only 3% trainable. It surpassed a size-matched variant of that baseline in every evaluated setting, with transfer gains increasing with shift severity: 2.1, 6.6, and 8.9 percentage points, respectively. Ablations attributed improvements to both pretraining and adapter recipe, with the pretraining contributing most under compound shift. Code and protocol are available at https://github.com/KeryanChelouche/udoppler-ssl.
Figures & tables
| Method | Trainable | Total | Glasgow | MAD |
|---|---|---|---|---|
| Full fine-tuning | ||||
| ResNet50 | 23.52 M (100%) | 23.52 M | 94.57 (1.08) | 97.94 (0.59) |
| DINOv3-S (full FT) | 21.60 M (100%) | 21.60 M | 94.04 (0.78) | 97.86 (0.61) |
| Linear probing on frozen models | ||||
| ViT-S (IN) | 0 (probe only) | 21.67 M | 87.16 (1.58) | 63.51 (1.24) |
| ViT-B (IN) | 0 (probe only) | 85.80 M | 88.26 (2.03) | 67.36 (1.88) |
| Method | Cross- age | Cross- orient. | Cross- dataset |
|---|---|---|---|
| Full fine-tuning | |||
| ResNet50 | 79.12 (0.32) | 79.37 (1.66) | 33.07 (3.71) |
| DINOv3-S (full FT) | 82.10 (0.73) | 81.35 (0.69) | 30.93 (2.97) |
| Linear probing on frozen models | |||
| ViT-S (IN) | 66.25 | 29.29 | 55.28 |
| DINOv3-S | 64.68 | 38.22 | 63.93 |
| Cross-orientation | Cross-dataset | |||
|---|---|---|---|---|
| ViT-S (IN) | DINOv3-S | ViT-S (IN) | DINOv3-S | |
| LoRA (attn) | 75.83 (0.83) | 80.49 (1.11) | 57.31 (0.92) | 65.54 (1.73) |
| PiSSA (all) | 81.19 (1.50) | 83.56 (1.02) | 58.15 (1.27) | 66.92 (0.38) |
Human Activity Recognition
Human activity recognition classifies actions and behaviors from wearable sensors, video, radio signals, and other measurements.
See all 242 papers