object-detection-fastai 0.0.3

Uses RetinaNet with FastAi

ObjectDetection

Some experiments with object detection in PyTorch and FastAi. This repo is created for educational reasons and to get a deeper understanding of RetinaNet and object detection general. If you like it, please let me know, if you find any bugs or tips for improvements also.

Install

The basline for this notebook was created by Sylvain Gugger from FastAi DevNotebook. Thank you very much, it was a great starting point and I'm a big fan off your work.

Library	version	link
Object-Detection-Metrics		https://github.com/rafaelpadilla/Object-Detection-Metrics

Publications using this code:

[x] Deep Learning-Based Quantification of Pulmonary Hemosiderophages in Cytology Slides

Examples:

• Medical images
• Coco subset

Results:

Features:

[x] Coco Metric at train time via callback [x] Flexibility

# use the feature map sizes 32,18,8,4 with 32 channels and two conv layers for detection and classification
RetinaNet(encoder, n_classes=data.train_ds.c, n_anchors=18, sizes=[32,16,8,4], chs=32, final_bias=-4., n_conv=2)
'''
  (classifier): Sequential(
    (0): Sequential(
      (0): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (1): ReLU()
    )
    (1): Sequential(
      (0): Conv2d(32, 32, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
      (1): ReLU()
    )
    (2): Conv2d(32, 18, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))
  )

# use the feature map sizes 32 with 8 channels and three conv layers for detection and classification
RetinaNet(encoder, n_classes=data.train_ds.c, n_anchors=3, sizes=[32], chs=8, final_bias=-4., n_conv=3)

[x] Debug anchor matches for training.

On the left image we see objects that are represented by anchors. On the right objects with no corresponding anchors for training. The size of the smallest anchors should be further decreased to match the small objects on the right image.