Multi-task Self-supervised Object Detection

This is an implementation of CVPR 2019: "Multi-task Self-supervised Object Detection via Recycling of Bounding Box Annotations" This is a novel object detection approach that takes advantage of both multi-task learning (MTL) and self-supervised learning (SSL). We propose a set of auxiliary tasks that help improve the accuracy of object detection.

The code is modified from Tensorflow Object Detection API.

Table of contents

• Preparation
  
• Exporting a trained model for inference
  
• Quick Start: Jupyter notebook for off-the-shelf inference
  
• Training and Evaluation
  

Evaluation of models

Model name (baseline/ours)	Detector	Backbone	Training	Eval	Baseline	Ours
model11 / model12	Faster R-CNN	ResNet101	VOC 07 trainval	VOC 07 test	77.0	78.7
model21 / model22	Faster R-CNN	ResNet101	COCO 2017 train	COCO 2017 val	32.7	34.6
model31 / model32	Faster R-CNN	ResNet101	VOC 07+12 trainval	VOC 07 test	81.7	83.7
model41 / model42	R-FCN	ResNet101	VOC 07 trainval	VOC 07 test	73.5	74.7
model51 / model52	Faster R-CNN	MobileNet	VOC 07 trainval	VOC 07 test	61.2	63.8
model61 / model62	Faster R-CNN	Inception ResNet v2	VOC 07 trainval	VOC 07 test	80.7	81.8
model71 / model72	R-FCN	ResNet101	VOC 07+12 trainval	VOC 07 test	78.6	80.6
model81 / model82	Faster R-CNN	MobileNet	VOC 07+12 trainval	VOC 07 test	68.6	70.8
model91 / model92	Faster R-CNN	Inception ResNet v2	VOC 07+12 trainval	VOC 07 test	84.3	86.0