Haoming Cai /haʊˈmɪŋ tsaɪ/

I am a thrid-year CS PhD student @ CS Department of University of Maryland, College Park, advised by Professor Christopher Metzler.

Previously, I had a wonderful journey (2020-2022) in X-Pixel with my supervisor Prof. Dong Chao (Shanghai AI Lab) and mentor Dr. Gu Jinjin (Univeristy of Sydney).

Research Interest - Computational Photography, Gen AI and Low-level Vision.

Contact : [email protected] / [email protected] / [email protected]

My CV : PDF

GitHub  /  Google Scholar  /  Flickr  / 

News

2025.02 - Two papers to appear in CVPR 2025. Congratulations to Tianfu, Mingyang, and Jingxi.

2025.01 - I will join Adobe NextCam team as 2025 summer intern.

Parametric Shadow Control for Portrait Generation in Text-to-Image Diffusion Models

Flash-Split: 2D Reflection Removal with Flash Cues and Latent Separation

TimeRewind: Rewinding Time with Image-and-Events Video Diffusion

Temporally Consistent Atmospheric Turbulence Mitigation with Neural Representations

Flash-Splat: 3D Reflection Removal with Flash Cues and Gaussian Splats

CodedEvents: Optimal Point-Spread-Function Engineering for 3D-Tracking with Event Cameras

This paper explores PSF engineering for neuromorphic event cameras, designing optimized masks for superior 3D point localization and tracking.

Snow Removal in Video: A New Dataset and A Novel Method

This paper presents a new deep learning framework for removing snow from videos, featuring a high-quality dataset and innovative modules for effective snow removal, outperforming existing methods.

Assessor360: Multi-sequence Network for Blind Omnidirectional Image Quality Assessment

Current omnidirectional image quality assessment lacks observer browsing modeling. We propose Assessor360, a novel multi-sequence network for BOIQA derived from realistic multi-assessor ODI quality assessment

Pipal: a large-scale image quality assessment dataset for perceptual image restoration

This paper highlights the challenge IQA faces with emerging GAN-based image restoration methods, noting a growing disparity between quantitative metrics and perceptual quality. To address this, the authors introduce a large-scale IQA dataset and benchmarks to enhance IQA methods’ effectiveness.

Super-resolution by predicting offsets: An ultra-efficient super-resolution network for rasterized images

SRPO is a real-time super-resolution method for computer graphics, achieving superior visual effects with minimal computational cost by leveraging rasterized image features and offset prediction.

Efficient image super-resolution using vast-receptive-field attention

This study improves super-resolution networks by refining the attention mechanism, leading to VapSR, which outperforms lightweight networks with fewer parameters, achieving similar results as IMDB and RFDN networks with significantly fewer parameters.

Blueprint separable residual network for efficient image super-resolution

Winner of Efficient Image Super-Resolution Track @ New Trends in Image Restoration and Enhancement (NTIRE) workshop of CVPR’23

Toward interactive modulation for photo-realistic image restoration

This paper presents CUGAN, a Controllable Unet GAN, for modulating image restoration tasks with fine texture details. Through dynamic level adjustments and condition networks, CUGAN outperforms previous methods, offering smooth user control over output effects.

Design and source code from Jon Barron's website