mvs-depth
ONNX export of a multiview-stereo depth model (dot-product cost-volume tier)
for the mvs package's live meshing pipeline.
depth_model_dot.onnx— opset 17, batch dynamic (1..N): current image Bx3x384x512, 7 source views, fp32 I/O with the fp16-autocast precision split baked into the graph (TensorRT 11 strongly-typed builds consume it directly).- Inputs:
cur_image_b3hw,src_image_bm3hw,src_K_bm44(matching scale s1),cur_invK_b44(s1),src_cam_T_world_bm44,cur_world_T_cam_b44. - Outputs:
depth_pred_s0_b1hw(Bx1x192x256 metric depth),lowest_cost_bhw. - Images: RGB / 255, ImageNet-normalized, bicubic-resized to 512x384.
Provenance: exported 2026-08-03 from the private inference repo's
tools/export_trt.py (fp16 autocast trace, dot_product checkpoint).
Weights derive from Niantic's SimpleRecon (CC BY-NC-SA 4.0); non-commercial.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support