This code is the implementation of paper -
Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks
PASCAL Visual Object Classes Challenge (2007) is used for training the network.
The dataset has 20 classes for detection and recognition task. For this network only 7 classes are used - aeroplane, bicycle, boat, bus, dog, train ,motorbike
There are ~2000 images pertaining to these classes. ~100 images are used for validation. One batch consists of a single image.
Data augmentation technique is used to reduce overfitting.
Network consists of
-
Anchors
- Corresponding to each image pixel in the convolution feature, there are 9 anchors at the scale of input image
- Corresponding to a 50*50 conv feature, there will be 50 * 50 *9 anchors at the image scale
- Anchors are used for training the region proposal and decider network
- All the predictions done are based on the anchors - e.g. if the anchor has object or Correction in the position of anchor
-
Region Proposal Network - generates ROI from input image
- Input image is resized to height (800) and width (800)
- Pretrained VGG16 model is used for generating Conv features. For VGG16, the size of each conv features is 50*50
- Conv features are passed through another conv layer, which further branches out to 2 output layers. One output layer decides if the corresponding anchor has object or not. Another calculates the correction in anchor position if the anchor has object.
-
Decider Network - proposes object classes from given ROI
- Corrected anchors from the region proposal network are passed through ROI pooling layer
- Output of ROI pooling layer is passed through decider network
- This network has 2 output layers. Class of ROI passed and the correction needed in ROI to accurately position the ROI around object.
Network is trained on google colab GPU.
A single image is considered as a training batch.
Average training loss per 20 iterations is calculated. Here is the graph -
After each epoch, network is validated on the validation dataset. Here is the graph -
There is a overfitting that can be seen in Net loss. This is due to overfitting in the RPN layer of the faster-rcnn network. Strangely, the increase of validation loss does not correspond to incorrect prediction. I believe, using detection validation schema similar to MAP can help in addressing this issue.
NMS is applied on the final predictions. Since, the network is not trained according to validation schema such as MAP optimization, this is not the optimized network.
