CVproject

Computer Vision project 2022

• Boldrin Cristian


• Bellan Riccardo

Models for Hand Detection

• First model: spec of EfficientDet-3, trained on whole EgoHandsDataset with 4800 images (3829 training, 476 validation, 495 test), 15 epochs and batch size 16
• Second model: transfer learning from yolov5s trained on whole EgoHandsDataset for 100 epochs, batch size 64
• Fine-tuned-model
• Augmented-model

Hand detection algorithm based on Yolov5 model for object detection

Hand detection algorithm based on Yolov5 model for object detection

Requirements:

• OpenCV >= 4.5.2
• GCC compiler

Methods for Hand Segmentation

• K-means
• Grabcut initialized with Rect
• Grabcut initialized with Mask (obtained by difference-from-skin computation)

Execution

CMake C++ Linux

mkdir build
cd build
cmake ..
cmake --build .
cd ..
./build/main path-to-image path-to-det-ground-truth.txt path-to-seg-ground-truth-mask.png

Displays and saves the results of the detection and segmentation process. It also saves a file ./output/det.txt containing bounding boxes coordinates and a file ./output/bin_mask.png containing the binary mask of the resulting Hand/NotHand segmentation.

Example of displayed image

./build/main input/07.jpg Results: