Hi. This is my implementation of the YOLO algorithm for object detection as a capstone project for my Computer Vision course.
We use a Cracker Box image dataset to train our model. For the purpose of conserving storage space, the dataset has been compressed and stored in the data.zip file, which contains 200 images of a cracker box in various positions and orientations in space. Each example consists of two (main) parts- <id>.jpg, a 640x480 image, and <id>-box.txt, a text file containing 4 values - <starting-x> <starting-y> <width of bounding box> <height of bounding box>. For visualization purposes, we also have a <id>-gt.jpg image, which is a 640x480 image with a bounding box around the cracker box.
| 000023.jpg | 000023-gt.jpg |
|---|---|
 |
 |
000023-box.txt
10.95 1.76 247.77 202.00
We use a Convolution Neural Network that I will not bother describing here, because it's 7 in the morning when I wrote this README, and I haven't slept all night. Anyways, the model architechture has been described here in the model.py file.
We train for 100 epochs, while using batch size of 8. We use half of our dataset for training and the other half for testing.
The training progress has been recorded below
We then use the trained model on the test dataset. We see that we achieve an average precision of 0.47.
Here is how our model performs with images from the test dataset, for which we provide no bounding box information.
The trained model (weights) is available on HuggingFace 🤗.
While the python files are present in the yolo directory, there is also a standalone Jupyter notebook YOLO.ipynb for your reading pleasure.
Thanks for your time 😃