W&B learning project: Bounding box object detection
Created on August 12|Last edited on October 7
Comment
Hi! For my W&B learning project I have opted to play around with object detection with bounding boxes using Keras, Tensorflow, and deep learning. I thought it would be cool to do some wildlife detection and stumbled upon a dataset of raccoon images with bounding boxes already labeled. Admittedly a dataset of around 200 images isn't large enough to produce best results but I just stuck with it since I really like raccoons.
Implementation
Bounding box object detection uses regression training as shown below to minimize loss. At the head of the network, we place a fully-connected layer with four neurons, corresponding to the bounding box (x, y) coordinates. Given that four-neuron layer, we implement a sigmoid activation function such that the outputs are returned in the range [0, 1]. The model then trains using a loss function such as mean-squared error or mean-absolute error on training data that consists of (1) the input images and (2) the bounding box of the object in the image. After training, we present an input image to our bounding box regressor network. Our network will then perform a forward pass and then actually predict the output bounding box coordinates of the object. (learning source)
Results
Very cool to see the bounding box predictions for every epoch trained
Misc Panels
Here is a sweeps run that conveniently takes a range of parameters for batch_size and number of epochs and runs all those experiments for you to try to get the lowest loss value.
Some neat stats on the GPU utilization, very useful for potential resource management
Add a comment