Faster RCNN

This is the Keras implementation of Faster RCNN architecture: Towards Real Time Object Detection with Region Proposal Network (RPN)

Note:

• Only tensorflow backend is supported and this requires tensorflow version 1.x
• Pydot and Graphviz packages need to be installed to visualize model architecture

Features:

• Tensorboard support for logging metrics and visualizing model graph
• Easy implementation for other feature extractors like VGG
• Terminal support for quick and easy execution

Usage:

Training:

• train_frcnn is used to train the model

        python train_frcnn.py -p sample.txt

• The input data should be a text file of training set with each line in the following format

  filepath,x1,y1,x2,y2,classname

  Example:

  /path/image_1.jpg,214,495,316,618,person

  /path/image_2.jpg,487,338,632,411,dog

• The model outputs the following:

  • Config pickle file – used during testing
  • Model weights in hdf5 format
  • A CSV file containing all the training metrics
  • Full model image for better understanding of the model architecture
  • Event file for visualizing model graph in tensorboard

Note:

• Input image size is set to 400. This can be changed in config.py.
• Pre-trained weights for ResNet can be downloaded from here

Testing:

• test_frcnn is used to test the model

        python test_frcnn.py -p test_images/ -a images.txt

• The input arguments should be a path to the test images folder and a text file with each line in the following format. The text file is needed to calculate mAP.

  filepath,x1,y1,x2,y2,classname

  Example:

  /path/image_1.jpg,214,495,316,618,person

  /path/image_2.jpg,487,338,632,411,dog

• The test file outputs the following:

  • Text files for each image with predicted bounding box coordinates, confidence score and its class (saved to a folder)
  • Test images with bounding box predictions (saved to a folder)
  • Prints the Mean Average Precision score
  • An image of Precison-Recall curve

Note:

• The index slicing should grab only the filename and not the file path in test_frcnn file. This is important for calculating mAP.

        with open(mean_ap, 'r') as f:
		for line in f:
			line_split = line.strip().split(',')
			(filename, x1, y1, x2, y2, classname) = line_split
			filename1 = filename[7:]
			if img_name == filename1:
				w = int(x2)-int(x1)
				h = int(y2)-int(y1)
				gt.append([str(filename1), int(x1), int(y1), w, h])
			else:
				pass

• Change the input image size in mean_avg file before testing

        gtFormat = BBFormat.XYWH
        detFormat = BBFormat.XYWH
        gtCoordType = CoordinatesType.Absolute
        detCoordType = CoordinatesType.Absolute
        imgSize = (800, 800)

API Reference:

https://github.com/tensorflow/models/tree/master/research/object_detection

References

• Implementation adopted from here and is based on the idea of tensorflow’s object detection API.
• Mean Average Precision implementation is from here