Wenzhao Zheng

I am currently a postdoctoral fellow in the Department of EECS at University of California, Berkeley, affiliated with Berkeley Artificial Intelligence Research Lab (BAIR) and Berkeley Deep Drive (BDD) , supervised by Prof. Kurt Keutzer . Prior to that, I received my Ph.D degree from the Department of Automation at Tsinghua University, advised by Prof. Jie Zhou and Prof. Jiwen Lu . In 2018, I received my BS degree from the Department of Physics, Tsinghua University.

I am interested in computer vision and deep learning. My current research focuses on:

Email  /  CV  /  Google Scholar  /  GitHub

Selected Publications

* indicates equal contribution

OccWorld models the joint evolutions of 3D scenes and ego movements and paves the way for interpretable end-to-end large driving models.

SelfOcc is the first self-supervised work that produces reasonable 3D occupancy for surround cameras.

As the first 2D-projection-based method on the 3D semantic occupancy prediction task, PointOcc significantly outperforms all other methods by a large margin with a much faster speed.

We design a pipeline to generate dense occupancy ground truths without expensive occupancy annotations, which enalbes the training of more dense 3D occupancy prediction models.

Given only surround-camera motorcycle RGB images barrier as inputs, our model (trained using trailer only sparse traffic cone LiDAR point supervision) can predict the semantic occupancy for all volumes in the 3D space.

We propose a unified framework for 3D perception and prediction based on multi-camera systems. The multi-task BEVerse outperforms existing single-task methods on 3D object detection, semantic map construction, and motion prediction.

We propose a SurroundDepth method to incorporate the information from multiple surrounding views to predict depth maps across cameras.

We propose an introspective deep metric learning (IDML) framework for uncertainty-aware comparisons of images.

We unify fully supervised and self-supervised contrastive learning and exploit both supervisions from labeled and unlabeled data for training.

We identify a token fluctuation phenomenon that has suppressed the potential of data mixing strategies for vision transformers. To adress this, we propose a token-label alignment (TL-Align) method to trace the correspondence between transformed tokens and the original tokens to maintain a label for each token.

This paper formulates deep metric learning under a unified framework and propose a dynamic constraint generator to produce adaptive composite constraints to train the metric towards good generalization.

We perform linear interpolation on embeddings to adaptively manipulate their hardness levels and generate corresponding label-preserving synthetics for recycled training.

We generate potential hard negatives adversarial to the learned metric as complements.

To address the ambiguity in attributions caused by different path choices, we introduced the Concentration Principle and developed SAMP, an efficient model-agnostic interpreter. By incorporating the infinitesimal constraint (IC) and momentum strategy (MS), SAMP provides superior interpretations.

This paper analyzes the consistency of existing Shapley value estimators and proposes the simple amortized estimator, SimSHAP. Extensive experiments conducted on tabular and image datasets validate the effectiveness of our SimSHAP, which significantly accelerates the computation of accurate Shapley values.

This paper proposes Bort, an optimizer for improving model explainability with boundedness and orthogonality constraints on model parameters, derived from the sufficient conditions of model comprehensibility and transparency.

This paper proposes an attributable visual similarity learning (AVSL) framework, which employs a generalized similarity learning paradigm to represent the similarity between two images with a graph for a more accurate and explainable similarity measure between images.

Website Template