Sandy3449

Job Title: Blender Python Developer & Technical Artist for Synthetic AI Training Data (Sim2Real) About the Project I am based in India and am building a scientific machine learning project involving synthetic training data. My work is focused on a Sim-to-Real pipeline for a pose-estimation AI. Project Overview This Phase 1 proof of concept will improve an existing Blender scene for a Sim-to-Real synthetic data pipeline. The scene contains a rigged 3D lab mouse, lab background footage, and camera calibration files. The work is to mathematically align the virtual scene, resolve a rig constraint issue, improve the photorealism, and automate the rendered images alongside 2D keypoint exports for pose-estimation training. Engagement & Phase 1 Constraints Timeline: The project must start on or before September 15, 2026, with delivery expected within 1 week of starting. Budget: €80–€100 fixed price milestone. Collaboration: This is Phase 1 (a proof-of-concept milestone) intended to support a longer-term collaboration. Animation and motion-data integration are outside this phase and are strictly reserved for Phase 2. What I Will Provide I will provide the existing .blend file, the calibration_cylinder.json file, real-lab GIFs/videos for reference, and screenshots detailing the coordinate-snapping issue privately after hiring due to lab privacy concerns. Technical Scope Scene Alignment and Rigging Use the calibration_cylinder.json containing intrinsic and extrinsic camera data to align the virtual Blender cameras with the real lab-camera space. Integrate two additional camera angles (e.g., top and bottom) using the calibration data. Diagnose and fix the rig coordinates/constraints that are snapping toward the world origin (0,0,0). This is likely a local/world space mismatch, an inverse-kinematics target offset, or a parenting issue. Verify the alignment against the supplied real-lab reference footage. Rendering and Dataset Automation Improve the mouse material and fur for photorealistic Cycles rendering. Add procedural or scripted variations for fur color, mouse scale, lighting, and environment. Write an embedded Blender Python script to render from the active cameras and export JSON or CSV files containing the exact 2D pixel coordinates for joints (nose, paws, base of tail, etc.). Use raycasting or equivalent visibility logic to mark or exclude keypoints that are occluded by the mouse body. Questions for Applicants Please answer these in your proposal so I know you have read the technical scope: Briefly explain how you would tackle the camera alignment using my JSON file. Looking at a scenario where rig constraints are jumping to the world origin (0,0,0), what are the first two things you would check in Blender? Which Blender Python function would you use to convert a 3D bone location into a 2D pixel coordinate for the export script?