Zhuqing Jiang

PBG-Net: Object detection with a multi-feature and iterative CNN model

We introduce PBG-Net, an object detection system based on an elaborately designed multi-feature deep CNN which works without proposal algorithms. Firstly, PBG-Net aggregates hierarchical features into multi-feature maps and discretizes the output of Conv5 feature map into a set of predicting boxes, namely Predicting Boxes Generation (PBG). Then, PBG-Net crops multi-feature maps via mapping the predicting boxes and handles the outcome into multi-feature concatenation. Finally, we exploit an iterative regression localization model based on a novel overlap loss function and online hard boxes selection. PBG-Net with around 100 boxes and an end-to-end joint training can achieve 74.2% and 71.1% mAP on the detection of PASCAL VOC 2007 and PASCAL VOC 2012 correspondingly at 12 fps on a NVIDIA GTX 1070p GPU, better than the Faster R-CNN counterparts.

Center contrastive loss regularized CNN for tracking

Visual tracking is a significant but challenging field in computer vision. Although considerable progress has been made in recent years, robust tracking in complicated scenes remains an open problem. Trackers get confused easily when similar objects appear or heavy clutter occurs due to indistinguishable features. In this work, a more effective feature extraction method based on convolutional neural network (CNN) is proposed. Different from conventional CNN models, this method applies a contrastive loss function to a single branch network, and centralization is employed to obtain more discriminative information. In addition, appropriate model update is used to capture transformation in object appearance. Quantitative experimental results on various video sequences demonstrate the superior performance of the proposed method in comparison with other state-of-the-art trackers. Complementary experiments are also conducted to validate some arguments.

Tracking with a small single Convolutional Neural Network

Robust visual tracking is a significant but challenging task in computer vision. Deep convolutional neural networks have been proverbially applied to visual tracking in recent years by learning a genetic representation from numerous training images. However, the deep networks training is time-consuming. In this work, an efficient and robust tracking algorithm using a small single Convolutional Neural Network (CNN) is proposed. Different from the existing CNN models, a novel loss function to process image batches in a single branch CNN is introduced. In addition, appropriate model update in a “when required” style is used when tracking to achieve performance boost. Quantitative experimental results on various video sequences demonstrate the superior performance of the proposed method in comparison with other state-of-the-art trackers.

Learning spatial-temporal consistent correlation filter for visual tracking

Discriminative correlation filters (DCF) have aroused great interests in visual object tracking in recent years due to the accuracy and computation efficiency. However, occlusion is still the main factor that affects performance. In this paper, a spatial-temporal consistent correlation filter utilizes the rich features extracted from a pre-trained convolutional neural network (CNN) is proposed to tackle this problem. We reformulate the conventional loss function and update classifier coefficients adaptively according to object appearance change rather than a constant learning rate. To acquire more accurate target location, this work combines correlation filter respond maps from different CNN layers together lie on their reliability. The experimental results evaluated on extensive challenging benchmark sequences demonstrate the proposed algorithm significantly improves the performance compared to state-of-the-art trackers.