Feature Map Activation Analysis for Object Key-Point Detection

Determining object location information in an image can enable more accurate CNN classification. Several extended CNN models have been developed to include both object location and classification into a unified model, at a cost of increasing compute complexity, increasing the number of parameters, and typically having a lower number of classification categories. We show that key-points within classifiable objects can be identified in early layer feature maps of a simple CNN without dependence on deeper layer processing or classification predictions. A statistical analysis of early feature maps is used to create a method for identifying and locating key-points that, with high probability, correspond to object locations in the image. This method uses only the forward pass of the simple CNN and requires no additional training. The method is tested on an image data set with known ground truth object locations as a function of the number of key points for four related selection methods. Results for object locations derived from key-points compare favorably to results obtained from R-CNN and are consistent over a range of key point set sizes.