Hi,
Thanks for making the dataset public.
I have a question about the data, and really hope you can provide help.
Since you just annotated the key-frames(samples) for both training, validation, and test, why do you provide the sweeps data? I mean, most of the data is not annotated, and it seems to be unnecessary to release the unannotated data. Or, we can use them in the other way? Thanks in advance!
Best
Hello,
In my case, it is really useful to have all the data at a higher framerate in order to use it in prediction and/or tracking tasks, since you can use, as @holger-motional told me in a past post, methods such as get_boxes().
Hope it helps.
Hi,
Thanks for your reply.
It does help.
You know, the dataset is a so large, but after I downloaded it, and studied the tutorial and the paper, I found the authors just annotated the key-frames, and results are only evaluated on the samples/key-frames. I doubted I misunderstood the data, and the evaluation requirements, so I asked this question. You explained clearly and helps me a lot. Thank you so much.
@ javierlorenzod
If you don’t want the sweeps you can chose not to download them.
But as @javierlorenzod says, they can be very helpful.
For the detection challenge we used up to 9 past sweeps to get temporal context.
Furthermore linear interpolation works very well for vehicles, so you essentially get many more “ground-truth” annotations.
Hi,
Thanks so much for your reply!
You and @javierlorenzod explained very clearly!
Best
@holger-motional