Missing masks

[tarashakhurana]

Hi, thanks for the amazing one-of-a-kind dataset! I downloaded and extracted the dataset but couldn't find the masks for this sequence train/cnb_dlab_0225_3rd. Could you verify on your end?

[y-zheng18]

Yes it seems the masks in that sequence are missing. Will update soon. Thanks!

[tarashakhurana]

Thanks for confirming! Additionally, I wanted to ask if there is a way to add semantic classes to the instance masks? For example, even if I am able to get the distinction between walls and other foreground objects, it'll be a huge win for me. Do you know of an easy way to get this?

[y-zheng18]

I'm sure it's possible since a mask color corresponds to a unique asset. You can refer to this code to see how to find the correspondences.

[m43]

Hej @y-zheng18, it seems that some of the datapoints do not have trajs_3d annotations but have array([None], dtype=object) instead. Do you happen to know why and if they could be added to the data? The following training data scenes are affected: character0_, character0_f2, character2_f, character5, gso_in_big.

[y-zheng18]

yes in those scenes the trajs_3d annotations are missing. We are fixing that and will update the dataset soon.

[saunair]

Hi @y-zheng18 I noticed the same issue with the trajs missing, do you have a version of the dataset where it's fixed?

[y-zheng18]

Hi @saunair, did you try the v1.2 data?

[saunair]

Hi @y-zheng18 Thanks for your reply!
Yes, this is with v1.2. I notice these scenes (not exclusvie), don't have the trajs3d key in them

character5_f
character6_f
character4_f
character3
character0_f2
character0_f
character0
character2
character5
character3_f
character1_f
character1

[saunair]

Hi @y-zheng18 I'm unsure if this is an issue caused due to extraction, as I only used the huge train.gz file in drive to get the train folders. When I look through the scenes I have, only see 104 scenes (I expected more). I don't see the issue of missing keys in val or test
I didn't use the .aa, .bb files, thinking the single big file would have all the necessary files. Is that assumption correct?

[y-zheng18]

It should be good. For character* sequences we don't have trajs3d annotations. To make up for that, in v1.2 we render more sequences like r0_new_, r0_new_f using similar assets.