LOVEU@CVPR'21 GEBD Sub-track 1.1 Forum

Go back to competition Back to thread list Post in this thread

> HMDB/ADL/UT-Ego Dataset Question

Hi organizer,

You released all the annotations of HMDB/ADL/UT-Ego Dataset. I'm not sure whether these data can be used in only task 1.1 or both.


Posted by: @ April 13, 2021, 3:17 a.m.

Hi Chen,

Thanks for your question. As stated on our white paper,

Sub-track 1.1:
- Can use any video data for training e.g. these videos not annotated in Kinetics, any web videos
- Can use supervision to train models for upstream tasks, e.g.:
object detection model trained on COCO, employ it on our Kinetics-GEBD to obtain intermediate results
action detection model trained on ActivityNet, employ it on our Kinetics-GEBD to obtain intermediate results
- In report, must specify what kind of external data or upstream supervision has been used.

Sub-track 1.2:
- Cannot use any videos NOT in our #Kinetics-GEBD#.
- Cannot use any additional supervision other than our GEBD annotations in Kinetics-GEBD. Thus, cannot use upstream task models that are trained on other dataset or annotation. (Exception: OK to use pre-trained backbone for initialization of your GEBD model)

Track 2:
(1) for our annotated #GEBD datasets (i.e. Kinetics, HMDB, ADL, UT-Ego), use 1 or multiple of them#
(2) do not use any of our annotations but use generic boundaries detected or annotated by yourselves on any dataset
(3) combination of (1) and (2) e.g. first train a GEBD model on our annotated datasets and then apply on another dataset to obtain large-scale pseudo boundary annotations for training a new model
Others - feel free to contact organizers to confirm if you are unsure

So you can use the recently released HMDB, ADL and UT-Ego GEBD annotations in Sub-track 1.1 and Track 2.


Posted by: leiwx52 @ April 13, 2021, 3:37 a.m.

Hi Stan,

track 1.1 and track 2, i get it, Thanks!


Posted by: @ April 13, 2021, 3:50 a.m.
Post in this thread