Custom dataset generates middle format - but unable to create valid NuScenes object

[AlexIlis]

I have custom data which is actually Kitti STyle annotated. To be compatible with Nuscenes in mmdeetction, I generated middle format that is needed for Nuscenes.

However for evaluation, DetectionEval from nuscnes-devkit expects a Nuscenes object for its purpose. But the challenge is that since I'm using custom data, it does not come with metadata that is present with NuScnenes dataset - attribute.json, ego_pose.json, map.json, sample_data.json, visibility.json, calibrated_sensor.json, instance.json sample.json, scene.json, category.json,log.json, sample_annotation.json, sensor.json

Is there an ideal way to make Custom dataset evaluate in NuScenes style?

[whyekit-motional]

@AlexIlis perhaps you could try dropping the irrelevant items in this list in the NuScenes object: https://github.com/nutonomy/nuscenes-devkit/blob/583dff71734d832b285d08f169b1a7f162b840c8/python-sdk/nuscenes/nuscenes.py#L59-L60

However, I do suspect you would need some of those items, e.g attribute.json, if you want to use DetectionEval off-the-shelf

Otherwise, you might have to dig into DetectionEval as well to see what needs to be commented out to suit your custom dataset

[AlexIlis]

Thanks much for you reply. My goal was to extract gt from infos.pkl but the challenge is that our prediction results are in the format {"sample_token": "", "translation": [], "size": [], "rotation": [], "velocity": [], "detection_name": "", "detection_score": , "attribute_name": null}. This seems to be the exact format that v1.0-trainval/sample_annotation.json file has, something I don't have on my custom dataset

Is there a way to generate sample.json, sample_annotation from infos.pkl file ?

[whyekit-motional]

I'm not sure how your infos.pkl file looks like, but you could probably write some sort of script that parses your file into the required format