Document origin of pretrained models provided in autoware.universe

[VRichardJP]

The autoware.universe perception modules contains many packages relying on CNNs:

• https://github.com/autowarefoundation/autoware.universe/tree/main/perception/lidar_apollo_instance_segmentation
• https://github.com/autowarefoundation/autoware.universe/tree/main/perception/lidar_centerpoint
• https://github.com/autowarefoundation/autoware.universe/tree/main/perception/traffic_light_ssd_fine_detector
• https://github.com/autowarefoundation/autoware.universe/tree/main/perception/tensorrt_yolo
• https://github.com/autowarefoundation/autoware.universe/tree/main/perception/traffic_light_classifier

For each, a pretrained model is hosted on https://awf.ml.dev.web.auto but not much information is given on how each was trained. For example, the following information would be interesting to have:

• dataset used for training. If trained on a private dataset, at least some information on the nature of the data (e.g. camera images from Japanese traffic lights, lidar data from urban traffic in Europe...)
• training/testing set size, number of objects by types...
• test accuracy

The point is to have a rough idea of what kind of performance one can expect from these pretrained models, and also what it would take to train new ones.

Would it be possible to gather such information and document it?

[yukkysaito]

@miursh @aohsato @eratostennis can you comment?

[miursh]

@VRichardJP
We can gather details of the data currently used and writing it to README may be easy way. Is that enough? or maybe this discussion is not only that?
From the perspective of continuous deployment, let me and TIER IV MLOps member discuss about this.
cc @eratostennis @yukke42

[VRichardJP]

Yes, I think a small summary in the README of each package would be enough. The information should help the user to understand the package use cases and its limits

[miursh]

working on autowarefoundation/autoware.universe#1087