OpenRobotLab / EmbodiedScan

[CVPR 2024 & NeurIPS 2024] EmbodiedScan: A Holistic Multi-Modal 3D Perception Suite Towards Embodied AI
https://tai-wang.github.io/embodiedscan/
Apache License 2.0
484 stars 37 forks source link

[Docs] Correct number of categories? #60

Open chanhee-luke opened 3 months ago

chanhee-luke commented 3 months ago

Branch

main branch https://mmdetection3d.readthedocs.io/en/latest/

📚 The doc issue

Hi! Thanks for the awesome work and the extensive annotations! :)

I have a question about the number of categories. I read in the paper that the total number of categories is over 760, but I only see 288 categories in the .pkl files for both train and validation. I assume that's why some objects have bounding boxes but are labeled as "object"? Are you planning to release annotations for more categories, or is this all?

Thanks!

-Luke

Suggest a potential alternative/fix

No response

Tai-Wang commented 3 months ago

You can also see in the paper's statistics that we take the 288 categories with the most objects (>10 instances per category) for the detection benchmark. The remaining categories have much fewer objects and form the long-tailed distribution. We may consider releasing all of them for potential usage of these annotations if you think they are valuable for your research. Looking forward to more feedback about this issue.

chanhee-luke commented 3 months ago

Hi Tai, it was nice talking to you in CVPR!

I hope you will consider releasing the full annotations, as having a long-tailed distribution is closer to real-life scenarios where this dataset will be most beneficial. I am mainly working on long-horizon task planning, and having comprehensive coverage of objects is important for my research interests. Thank you!

Tai-Wang commented 3 months ago

Hi Luke, glad to see you coming to try our dataset and repo!

Thanks for your suggestion, and we will consider releasing the full annotation with long-tailed categories in the next update. We will try to release the new ARKitScenes' annotations and MMScan's language annotations together with your mentioned modification in about one month.

Please, Xiaohan@mxh1999, have a look at this issue and remember this issue during the update.

chanhee-luke commented 3 months ago

Appreciate the quick response! I'll be looking forward to those :)