Partition Generation from Clustering Using Key Point Detection

Feature maps in CNNs are the main mechanism for extracting object features that are subsequently decoded to form a classification prediction. Object location in an image is an important and challenging factor to enable accurate CNN classification estimates. Several extended CNN models have been developed to include both object location and classification in a unified model. Extending the development of a method for extracting key points from early feature maps of a trained CNN model, we show that clustering techniques can be used to generate region predictions and bounding boxes. These bounding boxes are used to crop test images which are evaluated using a simple CNN model to predict the classification for the object in the predicted region. The results from cropping and simple CNN evaluation are compared against ground truth bounding box baseline and RCNN detection.