How to sample the 600 instances from Nuscenes?

[flymin]

Hi,

In your paper, you said "We sample 4 images from each scene, with 600 instances overall" for Nuscenes. However, there are only 500 tokens in data/interesting_image_tokens.txt and most of them come from sweeps rather than keyframes. Could you please provide more details about how to perform the evaluation on the Nuscenes?

Thanks.

[alexanderswerdlow]

Hi @flymin!

Please try this pickle file. It should load a python list with a simple pickle.load(). Please let me know if it works for you and if so, I'll upload it to the repo.

[flymin]

Hi @flymin!

Please try this pickle file. It should load a python list with a simple pickle.load(). Please let me know if it works for you and if so, I'll upload it to the repo.

Thanks! In this pickle file, there are 602 tokens for keyframes from the CAM_FRONT channel. Is it the subset you used for Table 1?

[alexanderswerdlow]

Yes, that was the subset used. 600 was a mistake. The code to generate the subset took every 10th keyframe and I think nuScenes validation isn't exactly 6k leading to the weird 602 number.

[flymin]

Yes, it makes sense. Thank you.