Explore Inria’s and Czech Institute of Informatics, Robotics and Cybernetics’ paper titled “MegaPose: 6D Pose Estimation of Novel Objects via Render & Compare”.

Our partners introduce MegaPose, a novel method designed to accurately estimate the position and orientation (pose) of objects in 3D space, even for objects it has not been trained on. During inference, MegaPose requires only a clear view and a 3D CAD model of the object.

The method introduces three significant contributions:

• A pose refinement technique that compares expected views of the object from different angles, improving accuracy.
• A coarse pose estimation approach that makes initial estimations and refines them based on the disparity between expected and observed views.
• Utilization of a large-scale synthetic dataset with diverse object shapes and appearances for training, enhancing the model's ability to generalize to new objects.

Evaluation on multiple datasets demonstrates that MegaPose achieves state-of-the-art performance, even surpassing methods trained specifically on the objects they are tested on. This suggests MegaPose's effectiveness in accurately recognizing and positioning novel objects.

Read the full publication here.

Find all AGIMUS publications here.