EmreGneri/video-depth-anything-small-offline

Video Depth Anything Small, offline temporal32 head ONNX export (154x266)

Offline (non-streaming) temporal head of Video Depth Anything Small, used by Spatial-Canvas to prepare depth for uploaded videos in the browser (32-frame windows, stride 22, overlap 10, as in the official inference). The image encoder is not included here: it is the same encoder_fp16.onnx as in EmreGneri/video-depth-anything-small-stream.

• Upstream code: https://github.com/DepthAnything/Video-Depth-Anything (commit 4f5ae23172ba60fd7bc11ef671cca678842c7072).
• Weights: depth-anything/Video-Depth-Anything-Small, video_depth_anything_vits.pth (Apache-2.0; checkpoint sha256 in offline-manifest.json). Only the Small model is Apache-2.0; Base/Large are CC-BY-NC and are NOT used here.
• Modification: the official temporal head (DPT head with motion modules) exported to ONNX for a fixed 32-frame window at 154x266, taking the four encoder feature maps ([32, 209, 384] each) and returning 32 depth maps; converted to fp16. The wrapper was checked against the official head, and the fp32/fp16 ONNX outputs against PyTorch (max absolute errors and file sha256 in offline-manifest.json). Exporter: scripts/export-vda-offline.py in Spatial-Canvas. No retraining.
• License: Apache-2.0, see LICENSE.