This project implements object detection in construction sites using the latest versions of the YOLOv10 and RT-DeTr architectures, leveraging PyTorch and the Ultralytics libraries. The models were trained on AWS SageMaker, utilizing a ml.g4dn.xlarge instance. This instance includes 1 NVIDIA T4 Tensor Core GPU with 16 GB GPU memory. The instance was configured with a 50 GB EBS volume to handle the dataset's size.
The dataset used in this project is a curated subset of the SODA: A large-scale open site object detection dataset for deep learning in construction, by Rui Duan, Hui Deng, Mao Tian, Yichuan Deng, and Jiarui Lin. The original dataset can be found here.
This curated, "bite-sized" dataset contains:
- Training set: 1000 images with labels.
- Testing set: 100 images with labels.
The subset has the following distribution of labeled objects, bounding boxes dimensions and center coordinates:
The images are in .jpg format, and the labels have been transformed to .txt files, following the format expected by the YOLOv10 and RT-DeTr models. The original .xml label files are also included.
The label files follow this format:
| label | x-coord | y-coord | width | height |
|---|---|---|---|---|
| 1 | 0.3229 | 0.8477 | 0.0333 | 0.1157 |
- label: The class label of the object (numeric).
- x-coord: Normalized x-coordinate of the bounding box.
- y-coord: Normalized y-coordinate of the bounding box.
- width: Normalized width of the bounding box.
- height: Normalized height of the bounding box.
The performance of YOLOv10 and RT-DeTr was evaluated using various metrics, with the results summarized below:
| Metric | YOLOv10-n | RT-DeTr-l |
|---|---|---|
| mAP50 | 0.82427 | 0.9073 |
| mAP50-95 | 0.51262 | 0.5854 |
| Precision | 0.76006 | 0.86107 |
| Recall | 0.78887 | 0.89708 |
| Training Time | 0.569 hours | 1.924 hours |
| Inference Speed | 2.3ms per image | 15.1ms per image |
The models are capable of object detection:
YOLOv10
DeTr
, and these are also capable of real-time detection. You can download and deploy the best checkpoints of these models from runs/detect/results_yolo10_L_pt/weights/best.pt and runs/detect/results_detr_pretrained/weights/best.pt
The best results for this model were achieved with the pretrained version yolov10n.pt:
The normalized confusion matrix for YOLOv10:
The best results for the DeTr were achieved with the pretrained version rtdetr-l.pt:
The normalized confusion matrix for RT-DeTr:
