Method

Point Sampling from 3D Shape Reconstructions

1. Use alpha shapes to reconstruct a 3D surface S from the exhaust condensation point cloud in the original point cloud
2. Uniformly sample N∈[100,1000] points from surface S
3. A new sample is obtained by assigning the reflection intensity to each point according to the nearest principle

Point Cloud Augmentation Strategy

1. Probabilistically placed in clean point cloud data in generating a large number of exhaust gas condensation point cloud samples:
   1. Pgas probability is generated in the rear center, right rear corner or left rear corner of the target
   2. Ptop probability is generated at the top of the target
   3. Paug total probability controls the variation of noise in the data
2. Convert each point to spherical coordinates, resample the point cloud using the same resolution as the sensor's parameter settings: solves an issue where applying point cloud augmentation would violate the dataset's sensor physics

Noise Robustness Loss

Introduce noise loss: IoU between the real frame of the exhaust gas condensation point cloud and the predicted target frame (reduce the noise in the predicted target frame - calculate the loss with the number of noise points in the frame:

Schematic of how it works:

Experiments

Comparison of vehicle classes on the DENSE test set:

Aug represents the model retrained using data augmentation and adding noise loss.Most of the test data does not include exhaust emissions, and other weather effects can side-effect the enhancement of the model.

Comparison of vehicle class accuracy on the DENSE-GAS test set:

Combined with a noise robustness loss, forces the network to learn to distinguish between vehicles and nearby noise points

Accuracy comparison on the DENSE test set following the TANet noise experiment:

Noise represents the amount of noise added in the target box

Ablation experiment:

PointRCNN's vehicle class accuracy on DENSE-GAS test set

1. No noise loss is applicable, no matter what kind of data augmentation is not conducive to network training
2. Using the proposed generation method is significantly better than adding random noise to the target box
   

The effect of noise loss weights on the model:

1. Too low is not conducive to the model paying attention to noise
2. Too high is easy to keep the model away from noise, but also the target position (only focus on noise)