Vision Hokusai — Self-supervised Vision Backbone for Edge Inference

Challenge

To dodge cloud-inference latency and privacy constraints, we needed sub-20ms inference on phones, Jetson Orin Nano, and Intel NUC. MobileNet/EfficientNet families plateau on downstream accuracy, and we wanted to avoid licensing constraints tied to large supervised ImageNet pre-training.

Solution

ViT-Base/14 served as the teacher under DINOv3 multi-crop + iBOT objectives. Knowledge transferred to MobileViTv3-XS via feature-alignment distillation (FitNet + Attention Transfer), then OpenVINO POT and CoreML INT8 quantization kept accuracy loss within 1%. Evaluation unified across LVIS, ADE20K, and iNaturalist.

Results

• MobileViTv3-XS + INT8 hits 18 ms/frame on Jetson Orin Nano
• ImageNet-1k Linear Probe at 76.2% (teacher ViT-Base reached 81.4%)
• ADE20K semantic segmentation mIoU of 42.1
• Model size 6.3 MB — 74% smaller than the 24.8 MB FP32 baseline