This is the summary of the paper Deep Generative Modeling of LiDAR Data.

The paper covers the niche problem of generative modeling on lidar scans. Often, the data through the sensor is corrupted or has some inconsistencies. This might lead to a tumultuous behaviour within the robot thus affecting the output. To address this problem, developing an intelligent system is an important mission.

Generative modeling has shown massive growth in images but it is still in its toddler stages compared to 3D and lidar data. Thus, the authors have devised to project the 3D lidar sensory data to a 2D grid. The spherical lidar data is converted to a 2D grid through a two-step process. Step 1, the N (x,y,z) points are clustered into H parts on the basis of the angle of elevation. Points with the same angle of elevation buckets into the same cluster. In the next step, the 360 degree plane is divided into W parts. This gives us a grid of H x W. Information of each cell is accumulated into a tensor of size 3 — H x W x 3. This data can be easily converted to the polar coordinate system. The devised technique is used on the KITTI dataset.

Now that the data is prepared, the baselines are developed on AtlasNet using the KITTI data without any preprocessing. Paper talks about conditional and unconditional generation. In conditional generation the data is slightly tampered with some noise, here Gaussian Noise, and then remodeled to reconstruct the input intelligently. Since the modeling is not based on sampling but instead on reconstruction, VAE and AE are used. Simultaneously, for unconditional generation, GANs are used.

Rigorous evaluation criteria for generative modeling is still an open problem. But here the authors have used the Earth Mover metric — optimal transport problem — and Chamfer distance. Moreso, for training parameters they have randomly selected learning rate, latent dimensions and batch size and experimented with 10 different configurations.

The proposed adversarial network can produce highly realistic images. Also VAE for lidar data provides a very robust solution to deal with noisy data.