Chenjie Cao 曹辰捷

Welcome! Chenjie Cao is at Alibaba DAMO Academy, working as a researcher of computer vision. He received his Ph.D. degree from Fudan University supervised by Prof. Yanwei Fu. His research interests focus on computer vision, which includes image inpainting, image synthesis, multi-view stereo, 3D reconstruction and generation, and feature matching.

Email  /  CV  /  Google Scholar  /  GitHub

* Researcher of DAMO Academy, Alibaba Group, 2024 - present.

* Post doc of Computer Science and Technology, Fudan University, 2024 - present.

* Ph.D of Statistics (Machine Learning track), Fudan University, 2020 - 2024.

* Algorithm engineer, PingAn OneConnect, 2019 - 2020.

* Master of Computer Science and Technology, East China University of Science and Technology (ECUST), 2016 - 2019.

* Bachelor of Computer Science and Technology, East China University of Science and Technology (ECUST), 2012 - 2016.

In this work, we propose MVInpainter, re-formulating the 3D editing as a multi-view 2D inpainting task. Specifically, MVInpainter inpaints multi-view images with the reference guidance rather than intractably generating an entirely novel view from scratch.

In this work, we present Animate3D, a novel framework for animating any static 3D model. We present a large-scale multi-view video dataset (MV-Video) to train the novel multi-view video diffusion model (MV-VDM).

In this paper, we propose a Variance alignment based 3D-2D Collaborative Denoising framework, dubbed VCD-Texture, to address texture synthesis.

This paper proposes an efficient, sparse-controlled video-to-4D framework named SC4D, that decouples motion and appearance to achieve superior video-to-4D generation.

In this study, we comprehensively explore multi-view feature priors from seven pretext visual tasks, comprising thirteen methods. Our main goal is to strengthen Neural Surface Reconstruction (NSR) training by considering a wide range of possibilities.

We formulate the reference-guided multi-view synthesis as an contextual inpainting issue, which is addressed by pre-existing attention modules of pre-trained Text-to-Image model through prompt tuning.

We employ a single diffusion generative model to address various sub-tasks with task-specific prompts.

We introduce MVSFormer++, a method that prudently maximizes the inherent characteristics of attention to enhance various components of the MVS pipeline.

Leveraging MAE priors to stabilize the inpainting of StableDiffusion. Moreover, we propose an effective augmentation strategy to eliminate the chromatic aberration in inpainting.

A Siamese network with a reciprocal loss is proposed to classify inliers/outliers for two-view matching.

Matching pairs located in spatially informative keypoints of both reference and target views enjoy better pose estimation. Thus a transformer-based cascaded matching model and a simple yet effective NMS filter are proposed.

The extension of CVPR work --ZITS. We further discuss the influence of various image priors on inpainting, and choice to use learning-based edges (L-Edge) as the new prior to enhance the meaningful structure recovery.

Improving the optical flow estimation through the image matching pre-text task.

We explore the usage of pre-trained ViTs for multi-view stereo, and further propose to unify both classification and regression-based depth predictions. MVSFormer achieves 1-st place in Tanks-and-Temple, and achieve 2-nd place in GigaMVS2022.

Improving the image inpainting with Masked AutoEncoder pre-training and attention-based restoration.

Learning structural priors (lines, edges) with transformers at first, then recovering textures with FFC based CNNs.

Firstly introducing wavelet prior for image inpainting. We also use axial-transformer to enhance the face/structural recovery.

The extension of P2M++, which makes P2M++ also work on SDF-based methods.

Leveraging GAN inversion and differentiable edge detector to achieve an effective face editing.

Propose the local autoregressive for local editing. The two-stream convolution is proposed to tackle information leakage.

We firstly introduce segment lines to improve the image inpainting.

A comprehensive Chinese NLP benchmark.

My early work about using ensemble learning to address imbalanced problems.

Improving the BERT pre-training with Sentence Insertion. SiBert won 1-st of CMRC2019.

My early work about using ensemble learning to address imbalanced problems.

My early work about using ensemble learning to address imbalanced problems.

My early work about using SVDD to address imbalanced problems.

Template