PROJECT TITLE :

Object Detection from Scratch with Deep Supervision

ABSTRACT:

We propose Deeply Supervised Object Detectors (DSOD) as an object detection framework that can be taught from the ground up in this research. Off-the-shelf models pre-trained on large-scale classification datasets like ImageNet and OpenImage are heavily used in recent improvements in object detection. However, because of the varied goal function and diverse distributions of object categories, transferring pre-trained models from classification to detection may result in learning bias. Techniques such as fine-tuning on detection tasks could help to some extent, but they aren't fundamental. It will also be more challenging to transfer these pre-trained models across disparate domains (e.g., from RGB to depth images). As a result, training object detectors from scratch is a better answer to these crucial difficulties, which justifies our proposed strategy. Previous attempts in this manner have largely failed because to a lack of training data and unsophisticated object detection backbone network architectures. We propose a set of design principles for building object detectors from the ground up in DSOD. Deep supervision, which is facilitated by layer-wise dense connections in both backbone networks and prediction layers, is one of the main ideas in learning excellent detectors from scratch. We build our DSOD based on the single-shot detection architecture after including various other principles (SSD). The PASCAL VOC 2007, 2012, and COCO datasets are used to test our technique. DSOD routinely outperforms state-of-the-art approaches that employ significantly more compact models. In particular, DSOD surpasses SSD on all three test while only requiring 1/2 of the parameters. We also found that with only 1/3 of the parameters, DSOD can get comparable/slightly better results than Mask RCNN [1] + FPN [2] (under similar input size) with no extra data or pre-trained models (under similar input size).