Author: Diego Alejandro Velazquez
Supervisors: Jordi Gonzàlez
Presentation time: 11:15
Virtual Room: 4.4 | Live presentation URL
This master thesis is based on DETR, a work presented by FAIR, which treats the object detection problem as a set prediction problem directly, resulting in an end-to-end fully differentiable pipeline. This approach does not require any domain specific prior to be fed into the model, and its performance is comparable to current state of the art methods. We evaluate the effectiveness of this new approach on the task of logo detection and work on improving its results on the detection of small objects, while keeping the model simple and fully differentiable. We obtain good results when compared to a strong Faster R-CNN baseline.