Dequan Wang

Ph.D. at UC Berkeley

Biography

Dequan Wang (王德泉) is a final year Computer Science Ph.D. at UC Berkeley, advised by Prof. Trevor Darrell.

Before coming to Bay Area, he graduated from School of Computer Science at Fudan University in June 2016, working with Prof. Yu-Gang Jiang, Prof. Zheng Zhang, and Prof. Xiangyang Xue.

Interests

Computer Vision
Machine Learning
Robotics

Education

PhD in Computer Science
University of California, Berkeley
BSc in Computer Science
Fudan University

Recent Papers

On-target Adaptation

As for most adaptation methods, most of the parameter updates for the model representation and the classifier are derived from the source and not the target. However, target accuracy is the goal, so we argue for optimizing as much as possible on the target data.

Dequan Wang, Shaoteng Liu, Sayna Ebrahimi, Evan Shelhamer, Trevor Darrell

Fighting Gradients with Gradients: Dynamic Defenses against Adversarial Attacks

Dent improves adversarial/robust accuracy (%) by more than 30 percent (relative) against AutoAttack on CIFAR-10 while preserving natural/clean accuracy. The static defenses alter training, while dent alters testing, and so this separation of concerns makes dent compatible with many existing models and defenses.

Dequan Wang, An Ju, Evan Shelhamer, David Wagner, Trevor Darrell

ActNN: Reducing Training Memory Footprint via 2-Bit Activation Compressed Training

ActNN is a PyTorch library for memory-efficient training. It reduces the training memory footprint by compressing the saved activations. ActNN is implemented as a collection of memory-saving layers. These layers have an identical interface to their PyTorch counterparts.

Jianfei Chen, Lianmin Zheng, Zhewei Yao, Dequan Wang, Ion Stoica, Michael W. Mahoney, Joseph E. Gonzalez

CoDeNet: Efficient Deployment of Input-Adaptive Object Detection on Embedded FPGAs

Co-design of a deformable convolution operation on FPGA with hardware-friendly modifications, showing up to 9.76× hardware speedup. Development of an efficient DNN model for object detection with co-designed input-adaptive deformable convolution that achieves 67.1 AP50 on Pascal VOC with 2.9 MB parameters. The model is 20.9× smaller but 10% more accurate than the Tiny-YOLO. Implementation of an FPGA accelerator to support the target neural network design that runs at 26 frames per second on Pascal VOC with 61.7 AP50.

Qijing Huang, Dequan Wang, Zheng Dong, Yizhao Gao, Yaohui Cai, Bichen Wu, Kurt Keutzer, John Wawrzynek

Tent: Fully Test-time Adaptation by Entropy Minimization

Tent equips a model to adapt itself to new and different data during testing ☀️ 🌧 ❄️. Tented models adapt online and batch-by-batch to reduce error on dataset shifts like corruptions, simulation-to-real discrepancies, and other differences between training and testing data.

Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, Trevor Darrell

Dynamic Scale Inference by Entropy Minimization

Dynamic receptive field scale is optimized according to the output at test time. We optimize receptive field scales and filter parameters to minimize the output entropy. This gives a modest refinement for training and testing at the same scale, and generalization improves for testing at different scales.

Dequan Wang, Evan Shelhamer, Bruno Olshausen, Trevor Darrell

Joint Monocular 3D Vehicle Detection and Tracking

We present a novel framework that jointly detects and tracks 3D vehicle bounding boxes. Our approach leverages 3D pose estimation to learn 2D patch association overtime and uses temporal information from tracking to obtain stable 3D estimation.

Hou-Ning Hu, Qi-Zhi Cai, Dequan Wang, Ji Lin, Min Sun, Philipp Krähenbühl, Trevor Darrell, Fisher Yu

Monocular Plan View Networks for Autonomous Driving

Monocular Plan View Networks use a monocular plan view image together with a first-person image to learn a deep driving policy. The plan view image is generated from the first-person image using 3D detection and re-projection.

Dequan Wang, Coline Devin, Qi-Zhi Cai, Philipp Krähenbühl, Trevor Darrell

Blurring the Line Between Structure and Learning to Optimize and Adapt Receptive Fields

We compose free-form filters and structured Gaussian filters by convolution to define a more general family of semi-structured filters than can be learned by either alone. Our composition makes receptive field scale, aspect, and orientation differentiable in a low-dimensional parameterization for efficient end-to-end learning.

Evan Shelhamer, Dequan Wang, Trevor Darrell

Objects as Points

We represent objects by a single point at their bounding box center. Other properties, such as object size, dimension, 3D extent, orientation, and pose are then regressed directly from image features at the center location.

Xingyi Zhou, Dequan Wang, Philipp Krähenbühl

Deep Object Centric Policies for Autonomous Driving

We propose an object-centric perception approach to deep control problems, and focus our experimentation on au- tonomous driving. Existing end-to-end models are holistic in nature; our approach augments policy learning with explicit representations that provide object-level attention.

Dequan Wang, Coline Devin, Qi-Zhi Cai, Fisher Yu, Trevor Darrell

Convolutional Neural Networks on non-uniform geometrical signals using Euclidean spectral transformation

We present a novel neural network layer that performs differentiable rasterization of arbitrary simplex-mesh-based geometrical signals (e.g., point clouds, line mesh, triangular mesh, tetrahedron mesh, polygon and polyhedron) of arbitrary dimensions. We further provide examples of incorporating the DDSL into neural networks for tasks such as polygonal image segmentation and neural shape optimization (for MNIST digits and airfoils).

Chiyu Jiang, Dequan Wang, Jingwei Huang, Philip Marcus, Matthias Nießner

Deep Layer Aggregation

Deep layer aggregation unifies semantic and spatial fusion to better capture what and where. Our aggregation architectures encompass and extend densely connected networks and feature pyramid networks with hierarchical and iterative skip connections that deepen the representation and refine resolution.

Fisher Yu, Dequan Wang, Evan Shelhamer, Trevor Darrell

VisDA: The Visual Domain Adaptation Challenge

It is well known that the success of machine learning methods on visual recognition tasks is highly dependent on access to large labeled datasets. Unfortunately, performance often drops significantly when the model is presented with data from a new deployment domain which it did not see in training, a problem known as dataset shift. The VisDA challenge aims to test domain adaptation methods’ ability to transfer source knowledge and adapt it to novel target domains.

Xingchao Peng, Ben Usman, Neela Kaushik, Judy Hoffman, Dequan Wang, Kate Saenko

FCNs in the Wild: Pixel-level Adversarial and Constraint-based Adaptation

While performance is improving for segmentation models trained and evaluated on the same data source, there has yet been limited research exploring the applicability of these models to new related domains. We propose the first unsupervised domain adaptation method for transferring semantic segmentation FCNs across image domains.

Judy Hoffman, Dequan Wang, Fisher Yu, Trevor Darrell

Multiple Granularity Descriptors for Fine-grained Categorization

A subordinate-level label carries with an implied hierarchy of labels, each corresponding to a level in the domain ontology. Following the assumption that domain experts distinguish finer classes with visually distinctive features, hierarchies thus have embedded and latent knowledge. The goal is to explore the rich semantic relationships among these extra labels.

Dequan Wang, Zhiqiang Shen, Jie Shao, Wei Zhang, Xiangyang Xue, Zheng Zhang

Weakly Supervised Semantic Segmentation for Social Images

Several social images and the associated noisy labels which may be correct, incorrect or missing. We learn a joint model to simultaneously segment and recognize visual concept in images.

Wei Zhang, Sheng Zeng, Dequan Wang, Xiangyang Xue