Kardiac-PulsR/PulsR-Depth

Depth Anything V2 Small — LiteRT (TFLite) int8, 518×518

A LiteRT/TFLite conversion of Depth Anything V2 Small (revision 5426e4f0f36572d16453bbda7a8389317b1bef99, Apache-2.0). It is used by the PulsR Android app for on-device depth parallax. Pictures are processed on the phone; this file is the only download.

File	depth_anything_v2_small_518_int8.tflite
Size	27,688,576 bytes
SHA-256	82ef5863d3ba7bddde1d8d1245f18c49b33d6f6c23df92505104924cf6970ccd
Input	float32 [1,518,518,3], NHWC, RGB, values 0..1. ImageNet normalisation is built into the graph.
Output	float32 [1,518,518]: relative inverse depth (larger = nearer). Normalise min/max per image.
Quantisation	dynamic range: per-channel int8 weights, fp32 activations
Ops	standard TFLite builtins only (no Flex or custom ops)

Conversion. litert-torch 0.9.4, ai-edge-quantizer 0.9.0, torch 2.13.0 (CPU).

Quality check. Compared with fp32 PyTorch on six varied photos, Pearson correlation is 0.99983 on average and 0.99948 at worst.

Licence and attribution

Apache-2.0. See LICENSE and NOTICE. Depth Anything V2 is by Lihe Yang et al. (HKU / TikTok); its DINOv2 encoder is by Meta Platforms. This file was modified by conversion and quantisation.

@article{depth_anything_v2,
  title={Depth Anything V2},
  author={Yang, Lihe and Kang, Bingyi and Huang, Zilong and Zhao, Zhen and Xu, Xiaogang and Feng, Jiashi and Zhao, Hengshuang},
  journal={arXiv:2406.09414},
  year={2024}
}