YOLOv1-TensorFlow2.0

This is a TensorFlow2.0 implementation of the YOLOv1 paper https://arxiv.org/abs/1506.02640, with the following changes,

• The feature exactor resembles the one mentioned in the YOLO9000 paper
• Input size is changed from 448x448 to 608x608
• The output stride is reduced from 64 to 32, to capture smaller objects
• Used 9 boxes per grid location, the paper uses 2. [doing this did not help much]

Video results

<img src="http://img.youtube.com/vi/9wjTtiVUXnE/0.jpg" alt="IMAGE ALT TEXT HERE" width="240" height="180" border="10" /> <a href="http://www.youtube.com/watch?feature=player_embedded&v=knSVWLZa_sU " target="_blank"><img src="http://img.youtube.com/vi/knSVWLZa_sU/0.jpg" alt="IMAGE ALT TEXT HERE" width="240" height="180" border="10" />

<img src="http://img.youtube.com/vi/ZfF9SYCjxU8/0.jpg" alt="IMAGE ALT TEXT HERE" width="240" height="180" border="10" /> <a href="http://www.youtube.com/watch?feature=player_embedded&v=H8feQqaTftg " target="_blank"><img src="http://img.youtube.com/vi/H8feQqaTftg/0.jpg" alt="IMAGE ALT TEXT HERE" width="240" height="180" border="10" />

Weights

trained weights

TODO

• [x] Change backbone to the one from YOLO9000
• [x] Use tf.distribute.MirroredStrategy for multi gpu training
• [x] Use tf.data.Dataset to implement the data input pipeline
• [ ] Add quantitative results
• [ ] Use learning_rate schedule during training
• [ ] Add horizonal flip in data input pipeline

Outputs

Training

• The model was trained on the Berkeley Deep Drive (BDD) dataset, which has 70,000 training images and 10000 validation images
• The model was trained for 200 epochs with a learning_rate of 5e-4 and a batch size of 24 [8 images per gpu] with heavy augmentations [brightness, saturation, contrast]
• No learning_rate schedule was followed (but was used by the authors).