vonet-compcars

VoNet for CompCars vehicle orientation

This repository contains a PyTorch reproduction of VoNet for five-class vehicle-orientation classification. The classes are front, rear, side, front-side, and rear-side.

The checkpoint selected by validation accuracy is from epoch 28 of a 30-epoch run:

• Validation accuracy: 94.5378%
• Validation loss: 0.1774
• Training accuracy at the selected epoch: 94.0809%
• Parameters: 396,949
• Input resolution: 227 x 227 RGB

This revised implementation follows the paper's module-level design, but the paper did not publish every layer width, so results and parameter count differ slightly from the original Caffe implementation.

Usage

Loading this custom architecture executes the Python files in this repository. Review them before setting trust_remote_code=True.

from PIL import Image
import torch
from transformers import AutoImageProcessor, AutoModelForImageClassification

repo_id = "samyou/vonet-compcars"
processor = AutoImageProcessor.from_pretrained(repo_id)
model = AutoModelForImageClassification.from_pretrained(
    repo_id,
    trust_remote_code=True,
).eval()

image = Image.open("car.jpg").convert("RGB")
inputs = processor(images=image, return_tensors="pt")
with torch.inference_mode():
    logits = model(**inputs).logits

predicted_id = logits.argmax(dim=-1).item()
print(model.config.id2label[predicted_id])

Training

• Dataset: CompCars vehicle viewpoint annotations
• Optimizer: SGD
• Epochs: 30
• Learning rate: 0.001
• Momentum: 0.9
• Weight decay: 0.0005
• Dropout: 0.5
• Initialization: Xavier uniform
• Augmentation: random horizontal flip
• Normalization: ImageNet mean and standard deviation

Detailed checkpoint metrics and arguments are in training_summary.json.

Intended use and limitations

The model predicts coarse vehicle orientation from a single RGB image. It does not distinguish left from right and may be unreliable under occlusion, unusual viewpoints, image-domain shift, or for vehicle designs poorly represented by CompCars. Validate it for your environment before safety-critical use. CompCars dataset terms remain applicable to uses involving the original dataset.

References

VoNet:

@inproceedings{you2016vonet,
  title={VoNet: Vehicle Orientation Classification Using Convolutional Neural Network},
  author={You, Ratanaksamrith and Kwon, Jang-Woo},
  booktitle={Proceedings of the 2nd International Conference on Communication and Information Processing},
  year={2016},
  doi={10.1145/3018009.3018045}
}

CompCars:

@inproceedings{yang2015compcars,
  title={A Large-Scale Car Dataset for Fine-Grained Categorization and Verification},
  author={Yang, Linjie and Luo, Ping and Loy, Chen Change and Tang, Xiaoou},
  booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
  year={2015}
}

Source implementation: https://github.com/samyou/vonet