Keypoint density-based region proposal for fine-grained object detection using regions with convolutional neural network features
JT Turner, Kalyan Moy Gupta, David W. Aha · 2016
Recent changes to the topology of regional convolutional neural networks (rCNN) have allowed them to obtain near real-time speeds in image detection. We propose a method for region proposal alternate to selective search which is used in the current state of the art object detection [3] and introduce the fine grained image datasets. In a maritime surveillance setting, it may be important to not only identify an object approaching your position but also know the type of vessel (e.g., a civilian fishing vessel or an enemy destroyer). Our region proposal technique Keypoint Density Region Proposal (KDRP) is able to achieve levels of performance that are not no worse than selective search at a very high level of significance, while only taking 49% of the time of selective search in the rCNN pipeline.