Object Detection using Vicreg and RetinaNet

Example 1

image1

For this project we aim at carrying out an object detection task with variable sized input. We first researched on recent state-of-the-art methods, and then performed our downstream task using VICreg (Variance-Invariance-Covariance Regularization) (Bardes et al., 2021) to pretrain our ResNet backbone and RetinaNet (Lin et al., 2017) to finetune on labeled images. Our team achieved a mAP score of 0.125 on validation dataset and we identified several ways to improve upon our current approach as well as proposed novel methods that we should try out.

My github project

My detailed report

Updated: