dronefreak/r3d-18-ucf101

R3D-18 for UCF-101 Action Recognition

Lightweight 3D ResNet-18 fine-tuned on UCF-101. Good baseline for action recognition with fast inference.

Model Details

• Model Type: 3D ResNet-18
• Base Model: R3D-18 pretrained on Kinetics-400
• Parameters: 33.2M
• Input: 16 frames @ 112×112 resolution
• Classes: 101 action categories

Performance

Metric	Value
Accuracy	83.80%
F1 Score	0.828
Precision	0.842

Comparison:

• Published R3D-18: 82.8%
• This model: 83.80% (+1.0%)

Usage

import torch
# Load from HuggingFace
from huggingface_hub import hf_hub_download
from torchvision.transforms import Compose, Resize, CenterCrop, Normalize, ToTensor
model_path = hf_hub_download(repo_id="dronefreak/r3d-18-ucf101", filename="r3d18-ufc101-split-1.pth")
model = torch.load(model_path)

# Prepare video (16 frames, C×T×H×W)
transform = Compose([
    Resize((128, 171)),
    CenterCrop(112),
    ToTensor(),
    Normalize(mean=[0.43216, 0.394666, 0.37645], 
              std=[0.22803, 0.22145, 0.216989])
])

# Inference
with torch.no_grad():
    output = model(video_tensor)
    prediction = output.argmax(dim=1)

Training

• Dataset: UCF-101 Split 1 (9,537 train / 3,783 test videos)
• Epochs: 100
• Batch Size: 32
• Optimizer: SGD (lr=0.001, momentum=0.9, weight_decay=1e-4)
• Augmentation: ColorJitter, RandomHorizontalFlip, RandomCrop

Use Cases

✅ Best for:

• Baseline comparisons
• Transfer learning starting point
• Educational purposes
• Fast prototyping

⚠️ Consider alternatives for:

• Maximum accuracy (use MC3-18: 87.05%)
• Real-time inference (use spatial models)

Comparison with MC3-18

Model	Accuracy	Speed	Use Case
R3D-18	83.80%	Fast	Baseline, prototyping
MC3-18	87.05%	Moderate	Best performance

Limitations

• Trained only on UCF-101 (limited to 101 action classes)
• Requires 16-frame clips (not suitable for real-time single-frame)
• Best performance on similar action types to UCF-101

Citation

@misc{r3d_18_ucf101,
  author = {Saumya Saksena},
  title = {R3D-18 for UCF-101 Action Recognition},
  year = {2024},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/dronefreak/r3d-18-ucf101}}
}

License

Apache-2.0