Inflated Multinomial Matching for Anchor-Free Object Detection

Sammanfattning: This thesis presents a novel matching strategy, Inflated Multinomial Matching, which enables training of anchor-free object detection models based on convolutional neural networks. An important aspect of detection models is the integral usage of anchor boxes, where an anchor box is a bounding box with a preset and constant location, size and shape in the image. The matching strategy presented removes the need for anchor boxes, where it instead utilizes the similarity scores between ground truths and predictions in a stochastic way which lets detection models obtain several independent submodels where each submodel specializes towards predicting objects of a certain size and shape. The resulting properties is essentially mimicking the main benefits of anchor boxes in an unsupervised way. The intended behavior of the matching strategy is confirmed through a number of indicators monitored throughout the training process. Finally, a full scale object detection model is trained with Inflated Multinomial Matching and example detection results are showcased.