Submitted to: BOP Challenge 2025

Training data: None

Onboarding data: The method performs model-free onboarding using the static onboarding sequences.

Notes:

Onboarding: The known camera poses and the estimated object poses relative to each camera are used to align the two onboarding sequences (upward and downward views) into a common coordinate frame. With these poses, the method performs feature detection on the masked object regions (SIFT), sequential feature matching, and point triangulation to obtain a sparse SfM point cloud of the object. The resulting camera poses and sparse point cloud serve as input for training a 3D Gaussian Splatting model for 14k iterations. Once trained, the model is used to render 642 template images from camera viewpoints distributed on an icosphere around the object. The entire onboarding stage requires, on average, 4 minutes and 30 seconds.

Inference: For inference, the method adopts a variant of the CNOS proposal stage, using Fast-SAM to generate object proposals. DINOv3 features are then extracted from these proposals, and the matching stage follows the CNOS pipeline closely.