This project implements a YOLO-style object detection model trained on KITTI-resolution images that were resized to suit smaller YOLO architectures. It follows a YOLOv1-like, anchor-free approach using a Darknet-53 backbone with transfer learning. Future improvements include integrating anchors for better bounding box quality.
The original dataset consists of images with a resolution of 1242×375, which is large for compact YOLO models. To make training feasible while preserving the aspect ratio, images were resized to 640×192.
A YOLOv1-style grid-based detection head is used without predefined anchors. The backbone is Darknet-53, initialized with pretrained weights for faster convergence.
-
Aspect-ratio–preserving image preprocessing Images resized from 1242×375 to 640×192.
-
YOLOv1-style, anchor-free detection head Grid-based classification and bounding box regression.
-
Darknet-53 backbone with transfer learning Stabilizes early training and improves convergence.
-
Accurate confidence predictions The model produces reliable confidence scores on test samples.
-
Planned anchor support Future iterations will integrate anchors to improve detection quality.
-
Data Preprocessing
- Resize images to 640×192.
- Convert labels to YOLOv1-compatible grid format.
-
Model Architecture
- Darknet-53 backbone (pretrained).
- Anchor-free YOLOv1-style detection head.
-
Training Setup
- Transfer learning enabled.
- YOLO-style composite loss (classification, confidence, localization).
-
Evaluation
- Stable confidence predictions.
- Reasonable bounding box performance for an anchor-free model.
- Add anchor-based predictions.
- Experiment with multi-scale training.
- Evaluate with metrics like mAP.
- Explore YOLOv3-style heads once anchors are integrated.
- Python
- PyTorch
- Darknet-53 pretrained weights
- KITTI (or similar) dataset
git clone <your-repo-url>
cd <repo>yolov1/yolov1.ipynb