Fast-ACVNet (fastforks)

Fast-ACVNet (Xu et al. 2022): MobileNetV2 features; attention weights from a correlation volume (1/8 group-wise, propagated to 1/4; Fast-ACVNet+: 1/4) select the top-24 disparities of a concatenation volume, then a 3-D hourglass, top-2 soft-argmax and superpixel upsampling (max disparity 192).

This repo is the model card only: it holds no weights. fastforks downloads the authors' weights from the links below and checks them against the SHA-256 in the model card. Code: https://github.com/kleinicke/fastforks. Kernels for the kernels backend: https://huggingface.co/fastforks/kernels. Original code: https://github.com/gangweix/Fast-ACVNet at 61d5375e1816.

Task: stereo. Tier: experimental. Outputs: disparity. Backends: portable, kernels.

Input: a rectified left / right image pair. Output: a disparity map (pixels, left view).

Variants

Weights are the authors' files (not hosted here). SHA-256 shortened to 12 hex digits; the full hashes are in card.toml.

variant	weights (author link)	SHA-256	weights licence	commercial use
fast-acvnet-generalization@1	Fast-ACVNet/generalization.ckpt	b9d06b9d3df3	MIT (assumed from code)	yes
fast-acvnet-kitti12@1	Fast-ACVNet/kitti_2012.ckpt	221b21107e60	MIT (assumed from code)	yes
fast-acvnet-kitti15@1	Fast-ACVNet/kitti_2015.ckpt	b7d8c609b8b2	MIT (assumed from code)	yes
fast-acvnet-plus-generalization@1	Fast-ACVNet+/generalization.ckpt	bd2f79680519	MIT (assumed from code)	yes
fast-acvnet-plus-kitti12@1	Fast-ACVNet+/kitti_2012.ckpt	e3ce45f2dc62	MIT (assumed from code)	yes
fast-acvnet-plus-kitti15@1	Fast-ACVNet+/kitti_2015.ckpt	bf4c8ef707eb	MIT (assumed from code)	yes
fast-acvnet-plus-sceneflow@1	Fast-ACVNet+/sceneflow.ckpt	1a1328828baf	MIT (assumed from code)	yes
fast-acvnet-sceneflow@1	Fast-ACVNet/sceneflow.ckpt	ee7ca89926f7	MIT (assumed from code)	yes

Licence

Original-code strings, recorded as they are (research-use package; check them before any other use). Code: MIT. Files: LICENSE.

Code licence: MIT; the authors do not state a separate weights licence.

Training data: Freiburg SceneFlow (FlyingThings3D, Driving, Monkaa).

Licence evidence: https://github.com/gangweix/Fast-ACVNet/blob/61d5375e1816f124f4417cea506eeaddac1504e4/LICENSE.md

Input size

any H x W >= 32, padded to a multiple of 32 Tested up to (portable, fp16, RTX 5060 Ti 16 GB): fast-acvnet-generalization 3000x4000 (5.0 GiB peak allocated, 7.9 reserved); fast-acvnet-kitti12 3000x4000 (5.0 GiB peak allocated, 7.9 reserved); fast-acvnet-kitti15 3000x4000 (5.0 GiB peak allocated, 7.9 reserved); fast-acvnet-plus-generalization 3000x4000 (4.9 GiB peak allocated, 7.9 reserved); fast-acvnet-plus-kitti12 3000x4000 (4.9 GiB peak allocated, 7.9 reserved); fast-acvnet-plus-kitti15 3000x4000 (4.9 GiB peak allocated, 7.9 reserved); fast-acvnet-plus-sceneflow 3000x4000 (4.9 GiB peak allocated, 7.9 reserved); fast-acvnet-sceneflow 3000x4000 (5.0 GiB peak allocated, 7.9 reserved)

Speed and memory

Stereo, Middlebury 2014 half, 994x1482

model	time (ms): original fp32	original	portable	kernels	kernels speed-up: vs original	vs portable	peak memory (MiB, allocated / reserved): original	portable	kernels
fast-acvnet-generalization	100.7	60.6	54.69	38.16	1.6x	1.4x	1421 / 1506	664 / 736	739 / 1042
fast-acvnet-plus-generalization	138.4	80.9	70.56	46.41	1.7x	1.5x	1441 / 1522	662 / 756	790 / 1120

Startup and first call

model	load (s): original	portable	kernels	first call (s): original	portable	kernels
fast-acvnet-generalization	0.24 (375x1242)	0.18	0.20†	1.73 (375x1242)	0.08†	0.04†
fast-acvnet-plus-generalization	0.23 (375x1242)	0.10	0.10†	0.04 (375x1242)	0.07†	0.05†

Fixed-size speed and peak memory, all models

model	mode	640x480	1280x720	1920x1080	12 MP 4000x3000
fast-acvnet-generalization	portable	13.9 ms / 137 / 208	31.1 ms / 402 / 664	78.3 ms / 890 / 1418	492.1 ms / 5077 / 8074
fast-acvnet-generalization	kernels	6.8 ms / 169 / 582	22.0 ms / 435 / 608	55.7 ms / 954 / 1372	357.3 ms / 5445 / 7618
fast-acvnet-kitti12	portable	14.0 ms / 137 / 208	31.1 ms / 402 / 664	78.2 ms / 890 / 1418	492.4 ms / 5077 / 8074
fast-acvnet-kitti12	kernels	6.8 ms / 169 / 582	22.0 ms / 435 / 608	55.6 ms / 954 / 1372	357.3 ms / 5445 / 7618
fast-acvnet-kitti15	portable	13.9 ms / 137 / 208	31.1 ms / 402 / 664	78.2 ms / 890 / 1418	492.4 ms / 5077 / 8074
fast-acvnet-kitti15	kernels	6.8 ms / 169 / 582	22.0 ms / 435 / 608	55.6 ms / 954 / 1372	357.3 ms / 5445 / 7618
fast-acvnet-plus-generalization	portable	14.3 ms / 136 / 208	38.0 ms / 401 / 664	99.4 ms / 887 / 1418	623.6 ms / 5054 / 8076
fast-acvnet-plus-generalization	kernels	7.8 ms / 169 / 546	25.3 ms / 457 / 598	66.2 ms / 1006 / 1226	422.5 ms / 5743 / 7342
fast-acvnet-plus-kitti12	portable	14.4 ms / 136 / 208	38.0 ms / 401 / 664	99.4 ms / 887 / 1418	623.3 ms / 5054 / 8076
fast-acvnet-plus-kitti12	kernels	7.7 ms / 169 / 546	25.3 ms / 457 / 598	66.2 ms / 1006 / 1226	422.4 ms / 5743 / 7342
fast-acvnet-plus-kitti15	portable	14.3 ms / 136 / 208	38.0 ms / 401 / 664	99.4 ms / 887 / 1418	623.7 ms / 5054 / 8076
fast-acvnet-plus-kitti15	kernels	7.8 ms / 169 / 546	25.3 ms / 457 / 598	66.2 ms / 1006 / 1226	422.8 ms / 5743 / 7342
fast-acvnet-plus-sceneflow	portable	14.3 ms / 136 / 208	38.0 ms / 401 / 664	99.4 ms / 887 / 1418	623.7 ms / 5054 / 8076
fast-acvnet-plus-sceneflow	kernels	7.8 ms / 169 / 546	25.3 ms / 457 / 598	66.1 ms / 1006 / 1226	422.7 ms / 5743 / 7342
fast-acvnet-sceneflow	portable	14.0 ms / 137 / 208	31.1 ms / 402 / 664	78.2 ms / 890 / 1418	492.4 ms / 5077 / 8074
fast-acvnet-sceneflow	kernels	6.8 ms / 169 / 582	22.0 ms / 435 / 608	55.6 ms / 954 / 1372	357.6 ms / 5445 / 7618

Fixed-size cells: time (ms) / peak allocated / peak reserved memory (MiB).

• fast-acvnet-kitti12, fast-acvnet-kitti15, fast-acvnet-sceneflow: same network and settings as fast-acvnet-generalization, measured once (rows not repeated).
• fast-acvnet-plus-kitti12, fast-acvnet-plus-kitti15, fast-acvnet-plus-sceneflow: same network and settings as fast-acvnet-plus-generalization, measured once (rows not repeated).

Hardware, protocol and caveats: benchmarks (RTX 5060 Ti 16 GB, sm_120). Original = the authors' code at the precision of portable (fp16, or (fp32) where fp16 is unsupported or not finite); original fp32 = its default fp32 run. Peak memory is the torch allocator's.

Accuracy

Parity with the original code

Every variant is checked against golden outputs of the original code at its pinned commit. CPU, float32: max |diff| <= 0.002 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.02 px and p99 <= 0.25 px.

Usage

from fastforks import stereo

model = stereo.load("fast-acvnet-generalization")  # portable PyTorch on the GPU if there is one, else the CPU
result = model("left.png", "right.png")
result.disparity  # [H, W] float32 tensor, input resolution
result.backend  # what ran, e.g. Backend(mode='portable', device='cuda:0', dtype='float16')

Install with pip install fastforks (see https://github.com/kleinicke/fastforks). Other variants: fast-acvnet-generalization, fast-acvnet-kitti12, fast-acvnet-kitti15, fast-acvnet-plus-generalization, fast-acvnet-plus-kitti12, fast-acvnet-plus-kitti15, fast-acvnet-plus-sceneflow, fast-acvnet-sceneflow.

Kernels backend

stereo.load("fast-acvnet-generalization", mode="kernels") runs our CUDA kernels (igev, cutlass_conv, cudnn_hacc). It needs a prebuilt kernel for your GPU architecture (downloaded from https://huggingface.co/fastforks/kernels on first use) and a recent NVIDIA driver; if there is none it raises BackendUnavailableError with the reason and never falls back silently. mode="auto" picks kernels when it can run and logs the choice.