Selected Work | Haixi Zhang

I design, train, and deploy neural perception systems. As technical owner of three ML/CV systems, I build the models, data pipelines, distributed training, and onboard inference that turn geometric learning into measurable results. My shared calibration-data pipeline streamed 17+ TB of raw logs in a single pass to produce ~600 GB of data for both calibration models.

Industry Systems

2production systems deployed

2 regionsautomated data collection

~13×training acceleration

SE(3) / SO(3)geometry-aware neural learning

01 · LCCNET · CAMERA ↔ LIDAR

Production · Deployed

Neural Camera–LiDAR Calibration

Designed, trained, and deployed a 6-DoF camera–LiDAR calibration network with dual encoders, camera-conditioned FiLM, and cross-modal attention. On an unseen vehicle, the deployed model reduced rotation error by 95% (2.95° → 0.14°) and reprojection error by 89% (22.4 → 2.4 px).

95%less rotation error · unseen vehicle

89%less reprojection error · unseen vehicle

~100Ktraining frames

~13×training acceleration

LCCNet neural network with dual encoders, camera-conditioned FiLM, cross-modal attention, and rotation/translation heads

NEURAL NETWORK DESIGNFiLM + cross-attention · click to enlarge

02 · LILINET · LIDAR ↔ LIDAR

Production · Deployed

Neural LiDAR–LiDAR Calibration

Built a LiDAR–LiDAR calibration network from scratch with shared spherical encoders, circular azimuth correlation, and sensor-conditioned rotation/translation heads. Deployed the model onboard, reducing fleet rotation error by 63% (0.93° → 0.35°).

63%less fleet rotation error

83%less rotation error · testing

96%evaluated frames improved

~20Ktraining frames

LiLiNet neural network with shared spherical encoders, circular azimuth correlation, sensor conditioning, and task-specific pose heads

GEOMETRIC NEURAL LEARNINGSpherical features + correlation · click to enlarge

03 · IN-CABIN · OCCUPANT PERCEPTION

Integrated · Vehicle Evaluated

In-Cabin Occupant Perception

Front- and rear-seat occupant monitoring built from RTMO detector/keypoint fine-tuning and a ResNet-18 multi-task occupancy/posture network, then integrated with iToF seating position, VLM temporal fusion, and confidence-aware vehicle messaging.

99.3%occupancy F1

95.8%posture Macro-F1

89.2%seat-depth accuracy

1.4 cmstable height MAE