The problem
State-of-the-art DPT models like Depth Anything V2 take ~10 seconds per frame on mobile devices (iPhone 13), making real-time or interactive on-device depth estimation impossible without drastic architectural compression.
Approach
Distilled a DPT teacher into a lightweight MobileViTV2-XS student equipped with a custom MiniDPT decoder. By combining SILog loss with gradient, feature, and self-attention distillation, the student preserves sharp depth edges while cutting inference latency from 10,000ms down to 170ms per frame on an iPhone 13.
At a glance
58x
Inference Speedup
170ms vs 10s on iPhone 13
92%
Accuracy Retention
57.3% vs 62.4% on δ < 1.25
SILog + 3 Distill Losses
Loss Functions
Feature, gradient & attention matching
ONNX → Flutter
Deployment
Android & iOS, on-device
Stack
PyTorchKnowledge DistillationMobileViTONNXFlutterDart