Pre-trained models for segmentation/classification

folira

Are there pre-trained networks available on Git currently that can be used for segmentation or object classification on OS-1 points_raw or images ? I did a research and I was just able to see some success driven through SuperPoint network which is actually for other purposes.
We would like to classify points corresponding to cars/humans/etc (or ultimately get to know their position) over OS-1 pointcloud, and as of now we don't have the ability to annotate/label our data to train a model. That's why we are looking for pre-trained models.

JakobGL

folira As described here, you can use YOLOv5 with Roboflow to label your data.

Mohit

Hi @folira
A computer vision company named Seoul Robotics offers solution for such applications. Also Ouster collaborated with the firm and the blog post about it can be found here.

mostafaeltalawy

Mohit
does ouster provide any documentation showing how to do object detection on ouster lidar data

JakobGL

mostafaeltalawy
Hi @mostafaeltalawy,
sorry for the late reply. This guide from the Ouster blog may be exactly what you are looking for. It explains how you can leverage pre-existing algorithms, originally designed for 2D image analysis, and adapt them to process 3D lidar data. The guide provides a step-by-step demonstration of this process.

Ouster also released Gemini, its own perception platform. Considering that you wanted to build the platform from scratch (as you wrote in your second post), YOLOv5 might be an interesting alternative to Gemini.

Best regards,
Jakob