August 2020
tl;dr: Use Domain Adaptation to bridge the gap between pseudo-lidar and real lidar.
DA-3Det uses a Siamese network and takes in real lidar and pseudo-lidar data. The difference between the features are penalized. This way DA-3Det learns a general feature based on pseudo-lidar.
Similar ideas to bridge the gap between real and pseudo-lidar has been witnessed in RefinedMPL, which proposes a way to downsample the dense lidar points to mimic the sparsity of point cloud.
- The paper also uses the Frustum PointNet version of pseudo-lidar due to its simplicity in dealing with point cloud.
- Siamese network with domain adaptation loss (L2 between features).
- During training process, real-lidar data is also utilized for feature domain adaptation. Only a single image is required during the inference stage.
- Context aware segmentation module: this is simply a pretrained segmentation module that is finetuned online.
- Pretraining improves performance as compared to unsupervised training with random initialization.
- Domain adaptation is a useful technique that can be applied to mono --> stereo and stereo --> lidar.
- Random sampling of lidar point for each object. For object containing smaller numbers of lidar points, sample with replacement (duplication).
- Questions and notes on how to improve/revise the current work