Sapiens-Pytorch-Inference

Minimal code and examples for inferencing Sapiens foundation human models in Pytorch

Why

• Make it easy to run the models by creating a SapiensPredictor class that allows to run multiple tasks simultaneously
• Add several examples to run the models on images, videos, and with a webcam in real-time.
• Download models automatically from HuggigFace if not available locally.
• Add a script for ONNX export. However, ONNX inference is not recommended due to the slow speed.
• Added Object Detection to allow the model to be run for each detected person. However, this mode is disabled as it produces the worst results.

Caution

• Use 1B models, since the accuracy of lower models is not good (especially for segmentation)
• Exported ONNX models are too slow.
• Input sizes other than 768x1024 don't produce good results.
• Running Sapiens models on a cropped person produces worse results, even if you crop a wider rectangle around the person.

Installation

pip install sapiens-inferece

Or, clone this repository:

git clone https://github.com/ibaiGorordo/Sapiens-Pytorch-Inference.git
cd Sapiens-Pytorch-Inference
pip install -r requirements.txt

Usage

import cv2
from imread_from_url import imread_from_url
from sapiens_inference import SapiensPredictor, SapiensConfig, SapiensDepthType, SapiensNormalType

# Load the model
config = SapiensConfig()
config.depth_type = SapiensDepthType.DEPTH_03B  # Disabled by default
config.normal_type = SapiensNormalType.NORMAL_1B  # Disabled by default
predictor = SapiensPredictor(config)

# Load the image
img = imread_from_url("https://github.com/ibaiGorordo/Sapiens-Pytorch-Inference/blob/assets/test2.png?raw=true")

# Estimate the maps
result = predictor(img)

cv2.namedWindow("Combined", cv2.WINDOW_NORMAL)
cv2.imshow("Combined", result)
cv2.waitKey(0)

SapiensPredictor

The SapiensPredictor class allows to run multiple tasks simultaneously. It has the following methods:

• SapiensPredictor(config: SapiensConfig) - Load the model with the specified configuration.
• __call__(img: np.ndarray) -> np.ndarray - Estimate the maps for the input image.

SapiensConfig

The SapiensConfig class allows to configure the model. It has the following attributes:

• dtype: torch.dtype - Data type to use. Default: torch.float32.
• device: torch.device - Device to use. Default: cuda if available, otherwise cpu.
• depth_type: SapiensDepthType - Depth model to use. Options: OFF, DEPTH_03B, DEPTH_06B, DEPTH_1B, DEPTH_2B. Default: OFF.
• normal_type: SapiensNormalType - Normal model to use. Options: OFF, NORMAL_03B, NORMAL_06B, NORMAL_1B, NORMAL_2B. Default: OFF.
• segmentation_type: SapiensSegmentationType - Segmentation model to use (Always enabled for the mask). Options: SEGMENTATION_03B, SEGMENTATION_06B, SEGMENTATION_1B. Default: SEGMENTATION_1B.
• detector_config: DetectorConfig - Configuration for the object detector. Default: {model_path: str = "models/yolov8m.pt", person_id: int = 0, confidence: float = 0.25}. Disabled as it produces worst results.
• minimum_person_height: float - Minimum height ratio of the person to detect. Default: 0.5f (50%). Not used if the object detector is disabled.

Examples

• Image Sapiens Predictor (Normal, Depth, Segmentation):

python image_predictor.py

• Video Sapiens Predictor (Normal, Depth, Segmentation): (https://youtu.be/hOyrnkQz1NE?si=jC76W7AY3zJnZhH4)

python video_predictor.py

• Webcam Sapiens Predictor (Normal, Depth, Segmentation):

python webcam_predictor.py

• Image Normal Estimation:

python image_normal_estimation.py

• Image Human Part Segmentation:

python image_segmentation.py

• Video Normal Estimation:

python video_normal_estimation.py

• Video Human Part Segmentation:

python video_segmentation.py

• Webcam Normal Estimation:

python webcam_normal_estimation.py

• Webcam Human Part Segmentation:

python webcam_segmentation.py

Export to ONNX

To export the model to ONNX, run the following script:

python export_onnx.py seg03b

The available models are seg03b, seg06b, seg1b, depth03b, depth06b, depth1b, depth2b, normal03b, normal06b, normal1b, normal2b.

Original Models

The original models are available at HuggingFace: https://huggingface.co/facebook/sapiens/tree/main/sapiens_lite_host

• License: Creative Commons Attribution-NonCommercial 4.0 International (https://github.com/facebookresearch/sapiens/blob/main/LICENSE)

References

• Sapiens: https://github.com/facebookresearch/sapiens
• Sapiens Lite: https://github.com/facebookresearch/sapiens/tree/main/lite
• HuggingFace Model: https://huggingface.co/facebook/sapiens