Shen Zheng 「郑深」

I am an first-year PhD student at the Robotics Institute, Carnegie Mellon University (CMU), working with Dr. Srinivasa Narasimhan . Previously, I worked as a Perception Software Engineer at Lucid Motors . I completed my Master of Science in Computer Vision (MSCV) degree at CMU. Prior to joining CMU, I earned my bachelor's degree in Math from Wenzhou-Kean University (WKU), where I worked with Dr. Gaurav Gupta .

Email1: shenzhen@andrew.cmu.edu

Email2: lebronshenzheng@gmail.com

CV  /  Google Scholar  /  Github  /  Linkedin  /  Leetcode  /  YouTube

My research focuses on improving scene understanding across diverse environments (e.g., weather and lighting) using the following strategies:

• Preprocessing:
  • Image Warping: Instance-Warp
  • Point Normalization: PointNorm
  • Low-Light Image Enhancement: SGZ, LLIE_Survey
  • Single Image Deraining: SAPNet
  • Blind Motion Deblurring: Deblur-YOLO
• Finetuning:
  • Image Generation: AS-IntroVAE
  • Image-to-Image Translation: TPSeNCE
  • Image Collection: ROADWork

Motivation: Domain adaptation methods struggle to learn smaller objects amidst dominant backgrounds with high cross-domain variations.

Solution: Oversample salient object regions by warping source-domain images in-place using instance-level saliency guidance during training while performing domain adaptation.

Motivation: Navigating through work zones is challenging due to a lack of large-scale open datasets.

Solution: Introduce the ROADWork dataset, which is so far the largest open-source work zone dataset, to help learn how to recognize, observe, analyze, and drive through work zones.

Motivation: Previous image-to-image translation methods produce artifacts and distortions, and lack control over the amount of rain generated.

Solution: Introduce a Triangular Probability Similarity (TPS) loss to minimize the artifacts and distortions during rain generation. Propose a Semantic Noise Contrastive Estimation (SeNCE) strategy to optimize the amounts of generated rain. Show that realistic rain generation benefits deraining and object detection in rain.

Motivation: Current point cloud analysis methods struggles with irregular (i.e., unevenly distributed) point clouds.

Solution: PointNorm, a point cloud analysis network with a DualNorm module (Point Normalization & Reverse Point Normalization) that leverages local mean and global standard deviation.

Motivation: Current low-light image enhancement methods cannot handle uneven illuminations, is computationally inefficient, and fail to preserve the semantic information.

Solution: Introduce SGZ, a zero-shot low-light image enhancement framework with pixel-wise light deficiency estimation, parameter-free recurrent image enhancement, and unsupervised semantic segmentation.

Motivation: Existing LLIE datasets focus on either overexposure or underexposure, not both, and usually feature minimally degraded images captured from static positions.

Solution: Present a comprehensive survey of low-light image enhancement (LLIE). Propose the SICE_Grad and SICE_Mix image datasets, which include images with both overexposure and underexposure. Introduce Night Wenzhou, a large-scale, high-resolution video dataset captured in fast motion with diverse illuminations and degradation.

Motivation: Generative models experience posterior collapse and vanishing gradient due to no effective metric for real-fake image evaluation.

Solution: Propose Adversarial Similarity Distance Introspective Variational Autoencoder (AS-IntroVAE), which can address the posterior collapse and the vanishing gradient problem in image generation in one go.

Motivation: Former deraining approaches often eliminate essential background details along with the rain, hindering tasks such as detection and segmentation.

Solution: SAPNet, an image-deraining network that integrates low-level image-deraining and high-level background segmentation using progressive dilated unit, perceptual contrastive loss, and unsupervised background segmentation.

Motivation: Previous unsupervised domain adaptation methods for medical imaging falter across varied imaging modalities due to substantial domain differences.

Solution: UDA-VAE++, an unsupervised domain adaptation framework that leverages mutual information maximization and sequential reparameterization for cardiac segmentation.

Motivation: Object detection algorithms exhibit suboptimal performance on blurry scenes.

Solution: Propose Deblur-YOLO, a generative adversarial network with a dilated feature pyramid generator, double multi-scale discriminators, and a detection discriminator to deal with photographs corrupted by motion blur.

Motivation: Existing ConvNets have poor computational complexity and require significant memory consumption.

Solution: Introduce an efficient ConvNet with weighted average stacking, Winograd-ReLU-based network pruning, and a electromagnetic-inspired dynamic batch size algorithm.

Director: Dr. Feng Guo

Working as a perception software engineer in the ADAS perception team responsible for traffic light detection, blockage detection, and battery fault detection.

Trained YOLO6 on full-resolution images containing traffic lights and fine-tuned arrow types, confidence, IoU, and area thresholds, resulting in a 40+% improvement in mAP (final mAP: 98%+ for day; 90%+ for night).

Developed a binary semantic segmentation model based on CenterNet to detect blockages such as ice, snow, mud, mud blur, rain drops, and sun glares, achieving 93%+ IoU.

Improved battery template matching to ensure success across all captured battery images, including in low-light or overexposed conditions.

Director: Dr. Wangjiang Zhu

Responsible for long-tailed data augmentation, training data auto-labeling and cleaning, and model evaluation for traffic light detection algorithms.

Implemented CycleGAN to conduct unsupervised data augmentation, converting traffic light bulbs from left arrow to round & leftUturn arrow.

Constructed a traffic light auto-label model using quantized VoVNet-57, filtering 14,618 incorrect annotations from 1,160,513 labeled frames.

Increased the classification accuracy for leftUturn traffic light from 78.41% to 87.27%, and the mean average precision from 93.01% to 94.80%.