Posts by Collection

portfolio

publications

Improving Device-Edge Cooperative Inference of Deep Learning via 2-Step Pruning

Published in Infocom workshop on IECOO, 2019

Wenqi Shi, Yunzhong Hou, Sheng Zhou, Zhisheng Niu, Yang Zhang, and Lu Geng

This paper proposes a novel 2-step pruning method that can drastically decrease the bandwitdh requirement for transimitting feature map from given layer. In such cases, the network can be partially deployed on mobile devices without increasing bandwidth requirement.

Vehicle Re-Identification with Location and Time Stamps

Published in CVPR workshop on AI-City, 2019

Kai Lv, Weijian Deng, Yunzhong Hou, Heming Du, Hao Sheng, Jianbin Jiao, Liang Zheng

In order to identify the similar looking vehicles, location and time stamps are used as cues in addition to appearance.

A Locality Aware City-Scale Multi-Camera Vehicle Tracking System

Published in CVPR workshop on AI-City, 2019

Yunzhong Hou, Heming Du, and Liang Zheng

Vehicle tracking across multiple cameras can be difficult for modern tracking systems. In order to avoid difficulties in a large scenario, we keep the tracking procedure within a minimal range.

Locality Aware Appearance Metric for Multi-Target Multi-Camera Tracking

Published in arXiv, 2019

Yunzhong Hou, Zhongdao Wang, Shengjin Wang, and Liang Zheng

We spot an inherent mismatch between local matching in multi-target multi-camera tracking (MTMCT) and global re-identification (re-ID) appearance feature. In fact, since targets move continuously, MTMCT only needs to match with a local region. To fit the local matching procedure in MTMCT, in this work, we introduce locality aware appearance metric (LAAM), that can be learned on top of several globally learned re-ID appearance features.

Learning to Structure an Image with Few Colors

Published in CVPR, 2020

Yunzhong Hou, Liang Zheng, Stephen Gould

We investigate compression for machine perception, the philosophy of which iscomparable to compression for human perception. Specifically, we restrict the color space to an extremely small size (only 1-bit image color), and then propose a CNN network to preserve the informative sturctures in an image. As opposed to the traditional clustering formulation, the proposed architecture, ColorCNN, formulates the color quantization problem as per-pixel classification. ColorCNN can be trained together with a classifier in an end-to-end manner. The quantization result from ColorCNN can achieve 82.1% accuracy with only 1-bit color on CIFAR-10, outperform the traditional quantization methods by a large margin.

Multiview Detection with Feature Perspective Transformation

Published in ECCV, 2020

Yunzhong Hou, Liang Zheng, Stephen Gould

We incorporate multiple camera views for pedestrian detection in heavy occluded scenes. Specifically, we propose an anchor-free fully convolutional multiview detector, MVDet, that relies on feature map perspective transformation. In addition, we create a novel synthetic dataset, MultiviewX, for additional evaluation.

talks

teaching

Teaching experience 1

Undergraduate course, University 1, Department, 2014

This is a description of a teaching experience. You can use markdown like any other post.

Teaching experience 2

Workshop, University 1, Department, 2015

This is a description of a teaching experience. You can use markdown like any other post.