Lingzhi Zhang

We propose a foundational panoramic scene understanding model capable of predicting multiple dense prediction tasks. We achieve this by i) curating large dataset with auto labeling pipeline by existing perspective prediction models, ii) propose PD-BridgeNet to tackle the multi-task interaction challenges under EPR distortions.

A foundational image soft effect removal model offering curated datasets, fine-grained user control via spatial masks, zero-shot generalization on unseen effects, and effect enhancement capabilities.

We generalize Perceptual Artifacts Localization to ten diverse image synthesis tasks, and show promising accuracy. We also show the effectiveness of automatic artifacts fixing and quality assessment as downstream applications.

We formulate a new task of learning perceptual artifacts localization on the inpainted images. We propose a high-quality labeled dataset, successfully train a model to localize the perceptual artifacts, and demonstrate downstream practical applications to no-reference single image quality assessment and iterative fill.

We propose a hybrid deep learning and patch-based approach for inpainting at modern camera resolution (4K+).

We propose a fine-grained egocentric hand-object segmentation dataset and model, and demonstrate its usage to multiple downstream vision applications, such as activity recognition, hand-object reconstruction, and seeing through the hands in videos.

We study the problem of common sense placement of visual objects in an image.

We proposed an image synthesis approach that provides stratified navigation in the latent code space.

We propose a deep optimization-based blending algorithm.

We study the problem of generating a set of realistic and diverse backgrounds when given only a small foreground region, which we formulate as image outpainting task.

Inspired by the properties of grid cells in mammalian brains, we build a generative model that enforces a normalized pairwise distance constraint between the latent space and output space to achieve data-efficient discovery of output spaces.