Diffusion-SS3D: Diffusion Model for Semi-supervised 3D Object Detection - Supplementary Material - A More Implementation Details

The proposed Diffusion-SS3D utilizes a teacher-student framework for 3D object detection in the setting of semi-supervised learning (SSL) and leverages the PointNet++ [3] as the encoder and the IoU-aware VoteNet [4] as the diffusion decoder. This section provides more details about the components in our implementation, including the encoder, decoder, diffusion initialization, SSL loss functions, and pseudo-label generation. Next, the noisy object size and noisy class label distributions are generated by adding Gaussian noise to the ground truth, and are used for this proposal box as the corrupted ground truth. Loss functions for object detection. In our experiments, we follow [4] and set λ to 2 for balancing the importance of labeled and unlabeled data.