Puzzles on LVIS

[Kewenjing1020]

Hi, as I know is that TAO contains sequences from 7 datasets, doesn't includes LVIS. But, why there's {train,validation}.json and {train,validation}_with_freeform.json in the annotation? Which one should I use for training? What's the relation between LVIS and TAO?

Besides, as you have mentioned, we can train the detector on LVIS-v0.5 and COCO. From the official website of LVIS, only v1..0 is accessible now. Could you please share the download link of LVIS-v0.5, or the mapping relation from v1.0 to v0.5?

[achalddave]

Thanks for the great questions!

• Why is there {train,validation}.json and {train,validation}_with_freeform.json? As noted in our paper, TAO contains objects labeled with LVIS categories, as well as free-form text categories. The {train,validation}.json files contain only the annotations for LVIS categories, while {train,validation}_with_freeform.json contain free-form text categories as well. You may use either of these for training. For the ECCV '20 challenge, we will only evaluate on the LVIS categories.

• LVIS v0.5: Yes, good point. I've uploaded the LVIS v0.5 annotations here: https://drive.google.com/drive/folders/1IHDbIvx0BQcj4BHkFLnEUwX4KOtaabTd?usp=sharing (alternative links v0.5 train, v0.5 val).