Automatic Annotation for Weakly Supervised Pedestrian Detection.

Pedestrian detection is an important task addressed in computer vision given its direct application in video surveillance, autonomous driving and biomechanics among many others. The advent of deep neural networks has meant a breakthrough in its resolution. The major problem is the need for very large labeled datasets, which is usually difficult to obtain, either because it is not publicly available or it is not suitable for the particular problem. To solve it, we design a method capable of self-labeling a detection dataset using only small manually labeled portion of it. Results show an autolabeled dataset of 10342 images from a preliminary set of 1312 manually labeled images.