This report outlines my approaches to fine-tune the YOLO-V7 model for detecting the largest and smallest objects in images and videos.

• There is no clear distinction between smallest and largest bounding boxes. Sizes alone cannot determine this distinction. • Some objects are exceptionally small (as small as 0.8 pixels), making accurate labeling difficult. They were removed from the data.

• Fine-tuning, then post-processing: I considered two variants: – (1) Fine-tune to detect largest and smallest boxes direct from the image, and then apply filtering to guarantee a single bounding box prediction per class. – (2) Fine-tune the model to detect objects in the image without distinguishing between specific objects, and then select two bounding boxes representing the smallest and largest sizes. – Model (2) yield better results in most cases.

Install the environment

conda env create -f environment.yml
conda activate amagcvTask1

Install and setup wandb as instructed here

Download the dataset and model and unzip to the desired folder

bash download.sh 

Two-class: Fine-tune the model to identify the largest and smallest object boxes within the image, then select the boxes that have the smallest and largest size. Single-class: Fine-tune the model to detect all objects in the image, then select two boxes that have smallest and largest size.

python utils/prepare_dataset.py --output_dir dataset/processed/ --min_size 8

The data used to trained and evaluated the models were uploaded here. Please download it and unzip to the dataset folder.

Two-class model

python train.py --workers 8 --batch-size 32 --data data/single_class_coco.yaml --img 640 640 --cfg cfg/training/yolov7_5anchors.yaml --weights downloaded_files/yolov7.pt  --name singleClass5anchor  --hyp data/hyp.scratch.p5.yaml --device 0

Single-class model

python train.py --workers 8 --batch-size 32 --data data/single_class_coco.yaml --img 640 640 --cfg cfg/training/yolov7_5anchors.yaml --weights downloaded_files/yolov7.pt  --name singleClass5anchor  --hyp data/hyp.scratch.p5.yaml --device 0

python test.py --batch-size 32 --data data/single_class_coco.yaml  --weights model.pt --device 0

Where model.pt is the path to your model

python detect.py --weight model.pt  --device 0 --img-size 640 --source inference/images/

• PDF Report: Link
• Wandb Report: Link