rtmo / README.md
Flor's picture
card: update from fastforks 1cbf1671
f8403b4 verified
|
Raw History Blame Contribute Delete
9.19 kB
metadata
library_name: fastforks
pipeline_tag: keypoint-detection
license: apache-2.0
tags:
  - pose
  - pose-estimation
  - human-pose-estimation
  - keypoint-detection
  - cuda
  - pytorch
  - fastforks
  - fastforks-card

RTMO (fastforks)

RTMO (Lu et al. 2024): real-time one-stage multi-person 2D pose from OpenMMLab. A YOLOX-style CSPDarknet backbone, a hybrid encoder neck (one transformer layer on the coarsest map) and a head that predicts, per grid cell, a person score, a box and a pose vector; a dynamic coordinate classifier turns each kept pose vector into the 17 COCO keypoints over 1D bins that span the person's box. Takes the whole image without person boxes or a separate detector: the backbone runs once per image, while decoding and NMS depend on the detections. The image is letterboxed to 640 x 640 (416 x 416 for rtmo-t); vendored from mmpose (no mmpose / mmcv dependency). The portable module() maps the letterboxed image to the head's per-cell outputs (decoding and NMS stay outside).

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/open-mmlab/mmpose at 5408bc76f5b8.

Task: pose. Tier: stable. Outputs: keypoints, keypoint_scores, boxes, scores. Backends: portable, kernels.

Input: one RGB image and a box per person. Output: keypoints with confidence scores for every person.

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
rtmo-l@1 rtmo-l-body7-640x640.pth b37118cee39b Apache-2.0 yes
rtmo-m@1 rtmo-m-body7-640x640.pth 39e78cc4afed Apache-2.0 yes
rtmo-s@1 rtmo-s-body7-640x640.pth dac2bf749bbf Apache-2.0 yes
rtmo-t@1 rtmo-t-body7-416x416.pth f48f75cbd59d Apache-2.0 yes

Licence

Original-code strings, recorded as they are (research-use package; check them before any other use). Code: Apache-2.0. Licence files: LICENSE.

Training data: COCO 2017 (train, instance annotations), AI Challenger human keypoints (2017), CrowdPose, MPII Human Pose, sub-JHMDB (J-HMDB), Halpe-FullBody, PoseTrack 2018 (body7: COCO, AI Challenger, CrowdPose, MPII, sub-JHMDB, Halpe and PoseTrack18 keypoints; the CSPDarknet backbone starts from a YOLOX COCO detection checkpoint).

Licence evidence: https://github.com/open-mmlab/mmpose/blob/5408bc76f5b848cf925a0d1857899011d8c5b497/README.md#license

Input size

any H x W >= 32, padded to a multiple of 32 Tested up to (portable, fp16, RTX 5060 Ti 16 GB): rtmo-l 3000x4000 (0.5 GiB peak allocated, 0.6 reserved); rtmo-m 3000x4000 (0.5 GiB peak allocated, 0.5 reserved); rtmo-s 3000x4000 (0.5 GiB peak allocated, 0.5 reserved); rtmo-t 3000x4000 (0.4 GiB peak allocated, 0.5 reserved)

Speed and memory

Pose, synthetic, 1080x1920

model time (ms): original fp32 original portable kernels kernels speed-up: vs original vs portable peak memory (MiB, allocated / reserved): original portable kernels
rtmo-l 17.1 21.4 10.8 8.30 2.1x 1.3x 800 / 824 208 / 226 208 / 228
rtmo-m 14.3 17.0 8.71 6.85 2.1x 1.3x 422 / 462 166 / 190 166 / 212
rtmo-s 11.4 15.5 7.48 6.27 1.8x 1.2x 228 / 252 142 / 160 142 / 180
rtmo-t 11.3 13.3 7.62 6.31 1.8x 1.2x 118 / 122 121 / 136 121 / 156

Startup and first call

model load (s): original portable kernels first call (s): original portable kernels
rtmo-l 0.55 0.36 1.52 0.06 0.24 0.16
rtmo-m 0.37 0.22 1.43 0.08 0.43 0.24
rtmo-s 0.27 0.12 1.37 0.08 0.41 0.26
rtmo-t 0.26 0.11 1.38 0.32 0.41 0.28

torch.compile and TensorRT

model input time: portable compile TensorRT fp16 kernels preparation: compile (cold first call) ONNX + engine build peak MiB (alloc / reserved): portable compile kernels TensorRT (engine + torch)
rtmo-l 1080x1920 10.8 ms 7.07 ms 2.90 ms (network on person crops only) 8.30 ms 18.9 s 39.6 s 208 / 226 192 / 224 208 / 228 277 / 306
rtmo-m 1080x1920 8.71 ms 4.99 ms 2.03 ms (network on person crops only) 6.85 ms 18.0 s 39.4 s 166 / 190 146 / 170 166 / 212 194 / 232
rtmo-s 1080x1920 7.48 ms 3.35 ms 1.21 ms (network on person crops only) 6.27 ms 14.6 s 36.7 s 142 / 160 121 / 156 142 / 180 133 / 164
rtmo-t 1080x1920 7.62 ms 3.30 ms 0.886 ms (network on person crops only) 6.31 ms 15.6 s 36.8 s 121 / 136 111 / 134 121 / 156 112 / 150

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

Hardware, protocol and caveats: benchmarks (RTX 5060 Ti 16 GB, sm_120). Original = the original code named above 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

On public datasets

Bottom-up on all 5,000 validation images, no ground-truth boxes; people ranked by their predicted person score, after model-specific decoding; pycocotools OKS AP. Bottom-up on the same every-tenth-image subset for all modes, ranked by predicted person score after model-specific decoding; pycocotools OKS AP. Original = the original code named under Details, same checkpoint and protocol.

variant dataset original fp32 original fp16 portable kernels note
rtmo-t COCO val2017, predicted people, keypoint AP 57.4 57.4 57.4 57.4
rtmo-t COCO val2017, every 10th image, predicted people, keypoint AP 59.4 59.4 59.3 59.3
rtmo-s COCO val2017, predicted people, keypoint AP 68.6 68.6 68.5 68.5
rtmo-s COCO val2017, every 10th image, predicted people, keypoint AP 71.1 71.2 71.2 71.2
rtmo-m COCO val2017, predicted people, keypoint AP 72.6 72.6 72.6 72.6
rtmo-m COCO val2017, every 10th image, predicted people, keypoint AP 74.2 74.2 74.2 74.2
rtmo-l COCO val2017, predicted people, keypoint AP 74.8 74.8 74.8 74.8
rtmo-l COCO val2017, every 10th image, predicted people, keypoint AP 76.1 76.0 76.0 76.0

Measured on NVIDIA GeForce RTX 5060 Ti, torch 2.14.1+cu130, fastforks be54b8b0, 2026-11-11.

Parity with the original code

Every variant is checked against reference outputs of the original code at its pinned commit. CPU, float32: max |diff| <= 0.014 vs the original. GPU, float16 (the default), portable and kernels: mean |diff| <= 0.47 and p99 <= 9.8 (absolute error of the output keypoints).

Limitations

  • module() exports the letterboxed-image network and raw per-cell head outputs; coordinate decoding and NMS remain outside the exported graph.
  • Float16 and float32 can retain different people near the score threshold.

Usage

from fastforks import pose

model = pose.load("rtmo-l")  # mode="auto": the preferred backend from recorded measurements
result = model("photo.jpg", boxes=[[x0, y0, x1, y1]])  # one box per person, or a fastforks.detection result
result.keypoints, result.keypoint_scores  # [P, K, 2] input pixels, [P, K]
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: rtmo-l, rtmo-m, rtmo-s, rtmo-t.

Kernels backend

pose.load("rtmo-l", mode="kernels") runs our CUDA kernels (cudnn_hacc, fvm_conv, fvm_extra, fvm_mb, fvm_r3). 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.