Model Card for Model ID

ViTPose: Simple Vision Transformer Baselines for Human Pose Estimation and ViTPose+: Vision Transformer Foundation Model for Generic Body Pose Estimation. It obtains 81.1 AP on MS COCO Keypoint test-dev set.

Model Details

Although no specific domain knowledge is considered in the design, plain vision transformers have shown excellent performance in visual recognition tasks. However, little effort has been made to reveal the potential of such simple structures for pose estimation tasks. In this paper, we show the surprisingly good capabilities of plain vision transformers for pose estimation from various aspects, namely simplicity in model structure, scalability in model size, flexibility in training paradigm, and transferability of knowledge between models, through a simple baseline model called ViTPose. Specifically, ViTPose employs plain and non-hierarchical vision transformers as backbones to extract features for a given person instance and a lightweight decoder for pose estimation. It can be scaled up from 100M to 1B parameters by taking the advantages of the scalable model capacity and high parallelism of transformers, setting a new Pareto front between throughput and performance. Besides, ViTPose is very flexible regarding the attention type, input resolution, pre-training and finetuning strategy, as well as dealing with multiple pose tasks. We also empirically demonstrate that the knowledge of large ViTPose models can be easily transferred to small ones via a simple knowledge token. Experimental results show that our basic ViTPose model outperforms representative methods on the challenging MS COCO Keypoint Detection benchmark, while the largest model sets a new state-of-the-art, i.e., 80.9 AP on the MS COCO test-dev set. The code and models are available at https://github.com/ViTAE-Transformer/ViTPose

Model Description

This is the model card of a 🤗 transformers model that has been pushed on the Hub. This model card has been automatically generated.

Developed by: Sangbum Choi and Niels Rogge
Funded by [optional]: ARC FL-170100117 and IH-180100002.
License: apache-2.0
Finetuned from model [optional]: [More Information Needed]

Model Sources [optional]

Repository: https://github.com/ViTAE-Transformer/ViTPose
Paper [optional]: https://arxiv.org/pdf/2204.12484
Demo [optional]: https://huggingface.co/spaces?sort=trending&search=vitpose

Uses

The ViTPose model, developed by the ViTAE-Transformer team, is primarily designed for pose estimation tasks. Here are some direct uses of the model:

Human Pose Estimation: The model can be used to estimate the poses of humans in images or videos. This involves identifying the locations of key body joints such as the head, shoulders, elbows, wrists, hips, knees, and ankles.

Action Recognition: By analyzing the poses over time, the model can help in recognizing various human actions and activities.

Surveillance: In security and surveillance applications, ViTPose can be used to monitor and analyze human behavior in public spaces or private premises.

Health and Fitness: The model can be utilized in fitness apps to track and analyze exercise poses, providing feedback on form and technique.

Gaming and Animation: ViTPose can be integrated into gaming and animation systems to create more realistic character movements and interactions.

Bias, Risks, and Limitations

In this paper, we propose a simple yet effective vision transformer baseline for pose estimation, i.e., ViTPose. Despite no elaborate designs in structure, ViTPose obtains SOTA performance on the MS COCO dataset. However, the potential of ViTPose is not fully explored with more advanced technologies, such as complex decoders or FPN structures, which may further improve the performance. Besides, although the ViTPose demonstrates exciting properties such as simplicity, scalability, flexibility, and transferability, more research efforts could be made, e.g., exploring the prompt-based tuning to demonstrate the flexibility of ViTPose further. In addition, we believe ViTPose can also be applied to other pose estimation datasets, e.g., animal pose estimation [47, 9, 45] and face keypoint detection [21, 6]. We leave them as the future work.

How to Get Started with the Model

Use the code below to get started with the model.

import numpy as np
import requests
import torch
from PIL import Image

from transformers import (
    RTDetrForObjectDetection,
    RTDetrImageProcessor,
    VitPoseConfig,
    VitPoseForPoseEstimation,
    VitPoseImageProcessor,
)


url = "http://images.cocodataset.org/val2017/000000000139.jpg"
image = Image.open(requests.get(url, stream=True).raw)

# Stage 1. Run Object Detector (User can replace this object_detector part)
person_image_processor = RTDetrImageProcessor.from_pretrained("PekingU/rtdetr_r50vd_coco_o365")
person_model = RTDetrForObjectDetection.from_pretrained("PekingU/rtdetr_r50vd_coco_o365")
inputs = person_image_processor(images=image, return_tensors="pt")

with torch.no_grad():
    outputs = person_model(**inputs)

results = person_image_processor.post_process_object_detection(
    outputs, target_sizes=torch.tensor([(image.height, image.width)]), threshold=0.3
)

def pascal_voc_to_coco(bboxes: np.ndarray) -> np.ndarray:
    """
    Converts bounding boxes from the Pascal VOC format to the COCO format.

    In other words, converts from (top_left_x, top_left_y, bottom_right_x, bottom_right_y) format
    to (top_left_x, top_left_y, width, height).

    Args:
        bboxes (`np.ndarray` of shape `(batch_size, 4)):
            Bounding boxes in Pascal VOC format.

    Returns:
        `np.ndarray` of shape `(batch_size, 4) in COCO format.
    """
    bboxes[:, 2] = bboxes[:, 2] - bboxes[:, 0]
    bboxes[:, 3] = bboxes[:, 3] - bboxes[:, 1]

    return bboxes

# Human label refers 0 index in COCO dataset
boxes = results[0]["boxes"][results[0]["labels"] == 0]
boxes = [pascal_voc_to_coco(boxes.cpu().numpy())]

# Stage 2. Run ViTPose
config = VitPoseConfig()
image_processor = VitPoseImageProcessor.from_pretrained("nielsr/vitpose-base-simple")
model = VitPoseForPoseEstimation.from_pretrained("nielsr/vitpose-base-simple")

pixel_values = image_processor(image, boxes=boxes, return_tensors="pt").pixel_values

with torch.no_grad():
    outputs = model(pixel_values)

pose_results = image_processor.post_process_pose_estimation(outputs, boxes=boxes)[0]

for pose_result in pose_results:
    print(pose_result)

Training Details

Training Data

Dataset details. We use MS COCO [28], AI Challenger [41], MPII [3], and CrowdPose [22] datasets for training and evaluation. OCHuman [54] dataset is only involved in the evaluation stage to measure the models’ performance in dealing with occluded people. The MS COCO dataset contains 118K images and 150K human instances with at most 17 keypoint annotations each instance for training. The dataset is under the CC-BY-4.0 license. MPII dataset is under the BSD license and contains 15K images and 22K human instances for training. There are at most 16 human keypoints for each instance annotated in this dataset. AI Challenger is much bigger and contains over 200K training images and 350 human instances, with at most 14 keypoints for each instance annotated. OCHuman contains human instances with heavy occlusion and is just used for val and test set, which includes 4K images and 8K instances.

Training Hyperparameters

Training regime:

Speeds, Sizes, Times [optional]

Evaluation

OCHuman val and test set. To evaluate the performance of human pose estimation models on the human instances with heavy occlusion, we test the ViTPose variants and representative models on the OCHuman val and test set with ground truth bounding boxes. We do not adopt extra human detectors since not all human instances are annotated in the OCHuman datasets, where the human detector will cause a lot of “false positive” bounding boxes and can not reflect the true ability of pose estimation models. Specifically, the decoder head of ViTPose corresponding to the MS COCO dataset is used, as the keypoint definitions are the same in MS COCO and OCHuman datasets.

MPII val set. We evaluate the performance of ViTPose and representative models on the MPII val set with the ground truth bounding boxes. Following the default settings of MPII, we use PCKh as metric for performance evaluation.

Results

Model Architecture and Objective

Hardware

The models are trained on 8 A100 GPUs based on the mmpose codebase [11]

Citation [optional]

BibTeX:

@misc{xu2022vitposesimplevisiontransformer, title={ViTPose: Simple Vision Transformer Baselines for Human Pose Estimation}, author={Yufei Xu and Jing Zhang and Qiming Zhang and Dacheng Tao}, year={2022}, eprint={2204.12484}, archivePrefix={arXiv}, primaryClass={cs.CV}, url={https://arxiv.org/abs/2204.12484}, }