I am a PhD Candidate at the University of Michigan, advised by Prof. Jia Deng and co-advised by Prof. David Fouhey. As a member of Princeton Vision and Learning Lab, I am doing research in computer vision and machine learning. My research focuses on exploiting compute graphics techniques for improving inverse rendering tasks such as 3D reconstruction, and adversarial attacks in 3D domains.
(April 2020) I will be joining Google!
(January 2020) Serving as a reviewer for ECCV 2020.
(June 2019) Serving as a Program Committee member of Security and Privacy of Machine Learning workshop at ICML 2019.
(May 2019) Serving as a reviewer for NeurIPS 2019.
(Mar 2019) Serving as a reviewer for ICCV 2019.
(Nov 2018) Serving as a reviewer for CVPR 2019.
Dawei Yang, Jia Deng
CVPR2020
Synthetic images rendered by graphics engines are a promising source for training deep networks. However, it is challenging to ensure that they can help train a network to perform well on real images, because a graphics-based generation pipeline requires numerous design decisions such as the selection of 3D shapes and the placement of the camera. In this work, we propose a new method that optimizes the generation of 3D training data based on what we call "hybrid gradient". We parametrize the design decisions as a real vector, and combine the approximate gradient and the analytical gradient to obtain the hybrid gradient of the network performance with respect to this vector. We evaluate our approach on the task of estimating surface normals from a single image. Experiments on standard benchmarks show that our approach can outperform the prior state of the art on optimizing the generation of 3D training data, particularly in terms of computational efficiency.
In this paper, we reveal the potential vulnerabilities of LiDAR-based autonomous driving detection systems. We test the generated adversarial objects on the Baidu Apollo autonomous driving platform and show that such physical systems are vulnerable to the proposed attacks. We also 3D-print our adversarial objects and perform physical experiments to illustrate that such vulnerability exists in the real world.
paper website
In this paper we consider adversarial behaviors in practical scenarios by manipulating the shape and texture of a given 3D mesh representation of an object. By generating 3D adversarial perturbation on shape or texture for a 3D mesh, the corresponding projected 2D instance can fool classifiers and object detectors.
paper oral poster bib
In this paper, we address the shape-from-shading problem by training deep networks with synthetic images. Our approach consists of two synergistic processes: the evolution of complex shapes from simple primitives, and the training of a deep network for shape-from-shading.
paper spotlight poster bib
In this work, we propose Decorrelated Batch Normalization (DBN), which not just centers and scales activations but whitens them. DBN retains the desirable qualities of Batch Normalization (BN) and further improves BN's optimization efficiency and generalization ability.
paper code bib
