What You’ll Do
Rapid SOTA Reproduction & Adaptation:
Implement training configs, hyperparameters, and losses/constraints (photometric reprojection, geometric consistency, uncertainty calibration).
On data, deliver train/val/test pipelines, A/B baselines, and reproducibility scripts; produce alignment and error reports (depth metrics, pose errors).
Design Training Data & Gains :
Design Training Dataset (sampling, rebalancing, synthesis, augmentation);
implement pre/post-processing to improve metrics (distortion correction, color consistency, depth fusion and hole filling, temporal consistency and uncertainty integration).
Job Requirements
Basic requirements:
Camera models of Equirectangular/cubemap, stitching/slicing, exposure/color alignment, seam/geometric consistency, stitching depth map.
Multi-view Geometry, triangulation/BA, depth fusion, uncertainty estimation; solid grasp of numerical optimization and linear algebra for vision.
Attention/positional encodings, cross-view & temporal attention, cost-volume and epipolar-aware constraints, long-context memory;
forward/reverse processes, noise schedules, conditional guidance, and sampling/distillation; with extensions or gains that preserve camera-model/epipolar consistency, scale observability, and uncertainty calibration.
Training & Evaluation & Coding Experience:
Proficient in Python (3.x)/PyTorch. Strong with NumPy/SciPy/OpenCV; able to deliver E2E training pipelines.
Familiarity with multi-GPU/distributed training, dataloader performance, mixed precision; able to export ONNX/TorchScript and validate accuracy/latency.
Solid C++ (C++17+) and libraries: Eigen/Ceres/g2o/OpenCV/Open3D (≥1).
Familiar with compilation tools such as CMake/Ninja, vcpkg/Conan.
Comfortable building/running/debugging in Docker (aligned with your duties on image size optimization and stability testing).
Fluent with and able to lightly extend COLMAP / OpenMVG / OpenMVS (≥1).
Preferred
Publications/Open-source: CVPR/ICCV/ECCV/NeurIPS/ICLR or notable repos.
Model-level contributions (architecture/loss changes, domain transfer, sparse-supervision utilization). Compensation uplift possible.
Experienced in optimizing model inference efficiency: TensorRT / OpenVINO / TVM.
Candidates with more than one year of relevant work experience (including internship experience) will be given priority consideration.

