dronefreak/human-action-classification
dronefreak/human-action-classification has 269 GitHub stars and 68 forks. GitHot measured 0 stars in 24 hours. Last checked 2026-11-18 18:55 UTC. Last push 2026-09-14.
Gemessener Sternzuwachs
- in 24h
- 0
- in 7 days
- Verlauf wird gesammelt
- in 15 days
- Verlauf wird gesammelt
- in 1 month
- Verlauf wird gesammelt
Nettoänderungen in überlappenden Zeiträumen, keine Tagessummen. Fehlende Messwerte bleiben leer. Sterne zeigen Interesse, nicht Qualität.
Über uns
Human action classification system with pose-based (MediaPipe) and video-based (3D CNN) models. Features 100+ architectures for real-time pose classification and temporal models pretrained on UCF-101/HMDB51.
Maschinelles LernenREADME.md
OriginalHuman Action Classification 🎬
Action recognition for both images and videos: pose-aware single-frame classification with MediaPipe + timm, and temporal 3D-CNN video classification with torchvision's video models.

Model Zoo
Video (UCF-101, 101 classes, 16-frame clips @ 112×112)
| Model | Accuracy | Params | Download |
|---|---|---|---|
| MC3-18 | 87.05% | 11.5M | |
| R3D-18 | 83.43% | 33.2M |
Video (HMDB51, 51 classes)
| Model | Init | Accuracy | Frames | Download |
|---|---|---|---|---|
| MC3-18 | Kinetics-400 | 56.34% | 8 | |
| MC3-18 | UCF-101 | 55.46% | 16 |
Both are reference baselines evaluated on split 1 only; see each model's card for the full training/evaluation writeup.
Image (Stanford40, 40 classes, single 224×224 frame)
| Model | Accuracy | Download |
|---|---|---|
| ResNet50 | 88.5% | |
| ResNet34 | 86.4% | |
| ResNet18 | 82.3% | |
| MobileNetV3-Large | 82.1% |
Also includes real-time pose classification (sitting/standing/lying) via MediaPipe, fused with the CNN prediction.
Full model list and architectures: timm_model_families.md.
Quick Start
pip install -e . # core library
pip install -e ".[demo]" # + Gradio web demo
pip install -e ".[dev,demo,train]" # everything
Video
from huggingface_hub import hf_hub_download
from hac.video.inference.predictor import VideoPredictor
model_path = hf_hub_download(repo_id="dronefreak/mc3-18-ucf101", filename="mc318-ufc101-split-1.pth")
predictor = VideoPredictor(model_path=model_path, device="cuda") # or "cpu"
result = predictor.predict_video("clip.mp4", top_k=3)
print(result["top_class"], result["top_confidence"])
CLI equivalent: hac-infer --video clip.mp4 --model <path> (add --num_frames 8 for the Kinetics-400-initialized HMDB51 model; UCF-101 and UCF-101-initialized HMDB51 models use the default of 16).
Image
from hac import ImagePredictor
predictor = ImagePredictor(model_path="weights/resnet50.pth", device="cuda", use_pose_estimation=True)
result = predictor.predict_image("person.jpg")
print(f"Pose: {result['pose']['class']}")
print(f"Action: {result['action']['top_class']} ({result['action']['top_confidence']:.2%})")
CLI equivalents: hac-infer --image photo.jpg --model weights/resnet50.pth, hac-infer --webcam --model ..., or hac-demo for the Gradio web UI (image pipeline only -- there is no video-model web demo yet).
More complete examples: scripts/simple_demo.py (annotated image output) and examples/quickstart.ipynb.
Training Your Own Models
# Video (UCF-101) -- after downloading UCF-101 + the official train/test splits
python -m hac.video.data.split_ucf101 --source UCF-101/ --output UCF-101-organized/ --splits ucfTrainTestlist/ --split_num 1
python -m hac.video.training.train --data_dir UCF-101-organized/ --model mc3_18 --pretrained --batch_size 32 --epochs 200
# Image (Stanford40)
python -m hac.image.training.train --data_dir data/ --model_name resnet50 --num_classes 40 --epochs 50 --batch_size 32
Both accept --help for the full set of options (mixup/cutmix, label smoothing, backbone freezing, resume, etc.). UCF-101: https://www.crcv.ucf.edu/data/UCF101.php. Stanford40: http://vision.stanford.edu/Datasets/40actions.html.
Contributing
Bug fixes, new architectures, mobile/ONNX deployment guides, and documentation are all welcome -- see CONTRIBUTING.md.
Citation & License
Apache-2.0 -- see LICENSE.
@software{saksena2026hac,
author = {Saksena, Saumya Kumaar},
title = {Human Action Classification: Image and Video Understanding},
year = {2026},
publisher = {GitHub},
url = {https://github.com/dronefreak/human-action-classification}
}
If you use the video models, please also cite the MC3/R3D architectures (Tran et al., 2018) and the datasets: UCF-101, Stanford40. Built with MediaPipe, timm, and PyTorch.