Need tips on improving performance on custom dataset

[Allamrahul]

Hi,
I am using this repo, specifically PointPillars for 3D object detection on custom point cloud dataset acquired using Ouster LIDAR. I am able to train the network and able to evaluate. My object of interest are farm posts in a barn. I used the ped_cycle xyres_16.config as reference since posts have a similar aspect ratio to a pedestrian and adapted it slightly. My objective is to detect four farm posts out of the many that are in front of the vehicle.

Currently, I have an eval of 30 point clouds. In that, I am able to get the detections in 75% of them. But in the others, its missing the posts. Not sure why that is. I have tried

• increasing batch size to 5
• tweaked the point cloud range
• tweaked the nms thresholds, anchor sizes
• Increased training data (520 point clouds)

BUT I am unable to achieve a performance greater than 75%.

The bev I is around 0.4. I need high recall i.e. getting all the 4 posts.

I would really like some suggestions in terms of what I can tweak in the config for better performance. I am attaching the config for reference.

xyres16.txt