Method About Sliding-window-based Weakly Labeled Object Detection

Object detection is an important branch of computer vision. Face detection in camera software, car and pedestrian detection in auto drive system, all these applications of object detection riches and will rich in all aspects of modern and future life, facilitating people with their cognition. Currently, object detection tasks generally rely on fully-labeled dataset such as Image Net and PASCAL VOC etc.These datasets not only offer the category information of the object in image, but also indicate its location and size in detail to facilitate the training of detector.However, in computer vision and its relevant field, most datasets are weakly-labeled,offering only information of whether object exists in the image, without detailed location and size information. Furthermore, fully-labeled datasets are man-made, as data scale grows, detailed information labeling will lead expensive time and money cost. Thus, compared with weakly-labeled datasets fully-labeled ones are rare resources.In view of former reasons, this project aims to achieve object detection tasks by using weakly-labeled datasets, giving prediction of both category and location information of the object in image.In this project we train and test on VOC2012 dataset. We treat VOC dataset as weakly-labeled during the training period. Model is first pre-trained on ILSVRC12 dataset, then transfers features from ILSVRC12 to VOC. We scan the input image by the use of sliding window strategy and output one score vector for each sliding window, indicating the score of this window on 20 VOC classes. Window with higher score implies higher probability for containing corresponding object. For each image, our model outputs high scored windows and their locations, regarding them as prediction of detected objects.This project defines metric for object detection task of sliding window strategy.It contains two different types with one for single-target detection and another for multi-target detection task. Under this metric, Weak Net model implemented for weakly-labeled detection in this project achieves nearly same prediction results with R-CNN model, which proves the effectiveness of our idea.