How to make gdino_kitti_base_oracle_2d.json/gdino_kitti_novel_oracle_2d.json/...?

[Hanqi7921]

Hi! I am currently meeting a problem. Right now I am doing an experiment with our Kitti dataset( (slightly modified from the original one) on your excellent evaluation work, I wonder how you make gdino_kitti_base_oracle_2d.json and gdino_kitti_novel_oracle_2d.json?

For example, in gdino_kitti_base_oracle_2d.json, I found there are images' id from 3712 to 3746, why you choose them? I found they are from the KITTI test, but not include all the KITTI test images. So why you just choose some of the images from KITTI test datasets as your gdino_kitti_base_oracle_2d? The same question goes to the file gdino_kitti_novel_oracle_2d.json.

It would be appreciated if you could share any idea or python script about how to use grounding dino to generate your gdino_kitti_base_oracle_2d.json and gdino_kitti_novel_oracle_2d.json. Thank uuuuuuu~

[yaojin17]

These files are from Grounding Dino’s inference results on kitti test dataset. These two files evaluate the base and novel categories separately. If an image in kitti test doesn’t contain the desired base or novel categories annotation, it is excluded from the file. So the base file contains almost all of the kitti test images. The novel file contains the images with the novel category "tram". The python scripts to generate these files are not cleaned up currently. It is not hard to write the scripts to get the Grounding Dino's inference results on kitti. If you encounter any problems, I am happy to provide assistance.

…

________________________________ From: Hanqi Lyu ***@***.***> Sent: Sunday, February 16, 2025 2:49:29 AM To: UVA-Computer-Vision-Lab/ovmono3d ***@***.***> Cc: Subscribed ***@***.***> Subject: [UVA-Computer-Vision-Lab/ovmono3d] How to make gdino_kitti_base_oracle_2d.json/gdino_kitti_novel_oracle_2d.json/...? (Issue #9) Hi! I am currently meeting a problem. Right now I am doing an experiment with our Kitti dataset( (slightly modified from the original one) on your excellent evaluation work, I wonder how you make gdino_kitti_base_oracle_2d.json and gdino_kitti_novel_oracle_2d.json? For example, in gdino_kitti_base_oracle_2d.json, I found there are images' id from 3712 to 3746, why you choose them? I found they are from the KITTI test, but not include all the KITTI test images. So why you just choose some of the images from KITTI test datasets as your gdino_kitti_base_oracle_2d? The same question goes to the file gdino_kitti_novel_oracle_2d.json. It would be appreciated if you could share any idea or python script about how to use grounding dino to generate your gdino_kitti_base_oracle_2d.json and gdino_kitti_novel_oracle_2d.json. Thank uuuuuuu~ — Reply to this email directly, view it on GitHub<#9>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/ANRSHE2HNA33HKPWEFEGPT32QA7ITAVCNFSM6AAAAABXHKD5TCVHI2DSMVQWIX3LMV43ASLTON2WKOZSHA2TKOJUGU2DINQ>. You are receiving this because you are subscribed to this thread.Message ID: ***@***.***> [Hanqi7921]Hanqi7921 created an issue (UVA-Computer-Vision-Lab/ovmono3d#9)<#9> Hi! I am currently meeting a problem. Right now I am doing an experiment with our Kitti dataset( (slightly modified from the original one) on your excellent evaluation work, I wonder how you make gdino_kitti_base_oracle_2d.json and gdino_kitti_novel_oracle_2d.json? For example, in gdino_kitti_base_oracle_2d.json, I found there are images' id from 3712 to 3746, why you choose them? I found they are from the KITTI test, but not include all the KITTI test images. So why you just choose some of the images from KITTI test datasets as your gdino_kitti_base_oracle_2d? The same question goes to the file gdino_kitti_novel_oracle_2d.json. It would be appreciated if you could share any idea or python script about how to use grounding dino to generate your gdino_kitti_base_oracle_2d.json and gdino_kitti_novel_oracle_2d.json. Thank uuuuuuu~ — Reply to this email directly, view it on GitHub<#9>, or unsubscribe<https://github.com/notifications/unsubscribe-auth/ANRSHE2HNA33HKPWEFEGPT32QA7ITAVCNFSM6AAAAABXHKD5TCVHI2DSMVQWIX3LMV43ASLTON2WKOZSHA2TKOJUGU2DINQ>. You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[Hanqi7921]

Sure, I'll try quickly! Thanks for your help~

[Hanqi7921]

Hello! I met one problem in making these gdino jsons.

My understanding about your Target-Aware Metrics is that you only provide words that existed in the pictures from the labels.txt as text_prompts in gdino 2d inference, instead of all the categories as text_prompts .

Is that the difference between gdino_kitti_base_oracle_2d.json and gdino_previous_eval_kitti_base_oracle_2d.json?

So, I try to make different text_prompts for each image. But it looks like the categories in the output->labels.json are not what I wanted. The error is that it can not split text_prompts correctly, for example, when I use text_prompts='car, van, cyclist', grounding dino may give out categories like 'car van', 'cyclist', instead of 'car', 'van', 'cyclist'. Do you meet this problem in ur work?

My language may not be clear, if you do not understand, you can ask further~

[yaojin17]

My understanding about your Target-Aware Metrics is that you only provide words that existed in the pictures from the labels.txt as text_prompts in gdino 2d inference, instead of all the categories as text_prompts .

Your understanding is correct.

Is that the difference between gdino_kitti_base_oracle_2d.json and gdino_previous_eval_kitti_base_oracle_2d.json?

"gdino_kitti_base_oracle_2d.json" is the result of using target-aware evaluation on base categories of kitti test. "gdino_previous_eval_kitti_base_oracle_2d.json" is the result of using previous standard evaluation on base categories of kitti test. Your understanding is also correct.

I didn't encounter this issue. For one object, grounding dino will output the probability of each token in the prompt. If the input is 'car, van, cyclist', it is likely that 'car, van' will be separated into different tokens. But there exist circumstances in which "motorcycle" may be divided into "motor" and "cycle" tokens. I provide some of our used code functions here. Hope these are useful for you!

[Hanqi7921]

I'll check it out. Thanks indeed!!!!!!! Have a great day^0^