valeoai / lost Goto Github PK

Is there a metric which indicates that there is no salient object in the image ? Im trying to apply LOST on a dataset in which the image can either have atleast one object, or no salient objects at all. Im wondering whether the degree of the seed patch can be a good metric to indicate whether or not there is an object in the frame.

Is it possible to share the scripts/config to train CAD on COCO? There's only one for VOC in the repo.
Thanks

Hi,

I just wanted to ask that when you gonna release the pre-trained model. Thanks

python main_lost.py --dataset COCO20k --set train

It will generate class-aware detection or class-unaware detection, or both?

I've replaced LOST's backbone (basically the dino weights) with the ones in CLIP, and it did not work well. But when switching back to dino weights, both ViT and ResNet50 backbone could generate good feature maps. Why would this happen?

Hi @osimeoni . I am suspicious about the corLoc evaluation part in the code. The corLoc for each image is true whenever one of the ground truth objects is hit!

Dear @osimeoni.
Could you release your trained Class-Agnostic Detector (CAD) models?
Thanks

Hi @osimeoni,

Thank you for making the code available!

When evaluating Detectron2 on VOC12 with the obtained pseudolables. I obtain the following error:
AttributeError: "int object has no attribute 'value'. It seems that the coco_style_file is not registered by 'register_coco_instances' (see image underneath). Any idea how this can be fixed? Thanks.

I’ve trained DINO’s model with my own Dataset, doing a finetuning on the ViT’s pre trained models of DINO. After a feel experiments I noticed that, every time that a epoch of the DINO’s finetune ran, the loss of the training reduce, however the IoU (the validation metric that we are using) of the bounding boxes generated by the LOST algorithm gets worse. Can anyone explain me why this is happening and how can I fix it?

HI, I have a confusion about the interesting work.
How to perform multi-target discovery in the figure 1 (middle) of your paper？
Any advice is greatly appreciated.

The predicted boxes are never assigned to preds_dict and therefore it just saves an empty dictionary

I am trying to run LOST + CAD on a set of unlabeled dataset I have. Is it possible to do that without any annotation?

Correct me if I am wrong, I need to get the LOST pseudo-boxes by running main_lost.py which should produce a .pkl file and results.text (which will be "nan" because I have no annotations, correct?)

After that I follow the instructions under "Training a Class-Agnostic Detector (CAD) with LOST pseudo-annotations". From what I read in the paper that this is fully unsupervised, but after trying to run the code and edit it to fit my dataset, I can see that GT annotation is needed?

Formatted my images into COCO style, and created an instance.json file for it, but in the annotation part there is only image ID and no segmentation.

TLDR;

I have unlabeled dataset, with no annotations, I want to run LOST+CAD fully unsupervised, is it possible?

In the image below is the message I get when I try to run "python main_lost.py --dataset DODO --set train --arch vit_small", DODO is the name of my dataset, I basically am using the COCO20k classes in the code but renamed and just changed the root_path.

Do you plan on releasing code for class-aware detection (i.e., to produce the results in Table 3 of https://arxiv.org/pdf/2109.14279.pdf)? I don't believe I see any of the necessary code for assigning object categories to boxes, but please correct me if I'm wrong.