You can deploy this model and start earning money today!

shijianjian/
ZDPShift

Name: ZDPShift
Availability: InStock

Views

No views yet

ZDPShift weights — Disparity Has a Sign

Checkpoints for Disparity Has a Sign: Stereo Matching Beyond the Zero-Disparity Plane. Each backbone comes as a pair:

Suffix	Meaning
`*_zs.pth`	the publicly released checkpoint, unchanged (zero-shot baseline; cannot output negative disparity)
`*_ours.pth`	the same backbone fine-tuned on horizontally translated SceneFlow only (no benchmark data, no new parameters); estimates signed disparity on both sides of the screen plane

Backbone	Files	Notes
FoundationStereo	`fs_zs.pth`, `fs_ours.pth`, `fs_cfg.yaml`	signed cost volume, `d_neg=64`, `d_pos=192`
IGEV-Stereo	`igev_zs.pth`, `igev_ours.pth`	signed cost volume, `d_neg=64`, `d_pos=192`
RAFT-Stereo	`raft_zs.pth`, `raft_ours.pth`	symmetric correlation range
StereoAnyVideo	`sav_zs.pth`, `sav_ours.pth`, `video_depth_anything_vits.pth`	the VDA file is the frozen monocular prior the model constructs
DynamicStereo	`dynamic_zs.pth`, `dynamic_ours.pth`
BiDAStereo	`bida_zs.pth`, `bida_ours.pth`

The ZDPShift benchmark itself is evaluation-only: none of these models has seen it. Code, evaluation scripts and the demo: https://github.com/shijianjian/ZDPShift · https://huggingface.co/spaces/shijianjian/ZDPShift · dataset: https://huggingface.co/datasets/shijianjian/ZDPShift

huggingface-cli download shijianjian/ZDPShift --local-dir weights