Point Cloud Segmentation Engineer

Point Cloud Segmentation Engineer | RGB-D | 3D Vision We're looking for an experienced Point Cloud Segmentation Engineer to solve a challenging industrial perception problem involving RGB-D sensing and 3D geometric reasoning This is not a conventional point cloud clustering or instance segmentation problem. The environment contains densely packed, highly occluded cuboidal objects where traditional Euclidean clustering fails because adjacent objects share faces and exhibit little or no visible separation. ### Problem Statement An RGB-D camera captures a scene containing multiple identical cuboidal boxes stacked tightly together on a pallet or inside a container. When the perception pipeline is triggered, the algorithm should: * Process the RGB-D image and generate a high-quality point cloud. * Analyze the complete 3D scene. * Detect every individual cuboid present in the stack. * Segment each cuboid as a separate instance, even when: * No physical gap exists between neighboring boxes. * Large portions of the boxes are occluded. * Only partial surfaces are visible. * All boxes lie on the same geometric plane. * Leverage 3D geometric reasoning rather than relying solely on Euclidean clustering. * Infer hidden object boundaries using: * Planar structures * Surface normals * Edge discontinuities * Shape priors * Known object dimensions * Spatial consistency ** Only one frame view will be there * Produce robust instance segmentation for every cuboid. * Estimate an accurate 6-DoF pose and Oriented Bounding Box (OBB) for each detected cuboid. * Deliver reliable performance under severe occlusion and noisy depth measurements. ### Input (Single Frame) * RGB image * Registered depth image * RGB-D point cloud * Camera calibration parameters * Known cuboid dimensions ### Expected Output For every detected cuboid: * Instance segmentation * Segmented point cloud * Oriented Bounding Box (OBB) * 6-DoF pose * Confidence score ### Required Skills * Point Cloud Processing (Open3D / PCL) * RGB-D Computer Vision & 3D Geometry * 3D Instance Segmentation & Geometric Reasoning * Pose Estimation & OBB Fitting ### Nice to Have * PointNet++, Point Transformer, Mask3D, SuperPoint Transformer * Shape completion * Plane extraction and geometric primitive fitting * CUDA / GPU acceleration ### The Challenge The primary challenge is that the cuboids are tightly packed with almost no visible separation. In many cases, adjacent cuboids share complete faces, making traditional clustering and segmentation approaches ineffective. The ideal solution should combine 3D geometry, shape priors, spatial reasoning, and robust perception algorithms to accurately infer object boundaries that are not explicitly visible in the raw point cloud. If you're passionate about solving complex real-world perception problems, we'd love to hear from you.