facebookresearch / vlpart Goto Github PK

thanks for your excellent work,i'm new to deep learning,
how to visualize the results of paco dataset, i used the swinbase_cascade_paco.yaml model and the checkpoints to train,but how to visualize the obtained json file? thanks a lot

Hello!

Thank you for this great work! I was wondering if there's an easy way that I'm missing to dynamically change the custom vocabulary per sample? Or is this a limitation of detectron2 (i.e., not being able to change Metadata)?

Hello!

I am trying to reproduce the pascal part results when going from base to novel categories and am experiencing different results from saving my own weights vs. using the ones you have provided in datasets/metadata.

For example, when using the saved weights you provide these are my results after 2000 iterations (1 gpu).

And these are my results after saving new weights using tools/pascal_part_base_clip_name.py.

The only difference seems to be in the precision between these weights (in the 1e-4 place). Do you know why there would be such a discrepancy?

Thank you!
Josh

Hi, I found this part of the code confusing. Could you give some explanations? I put some comments that I didn't understand. Thanks.

    for ann in data['annotations']:
        segs = ann['segmentation']
        new_segs = []
        for seg in segs:
            assert len(seg) > 0 and len(seg) % 2 == 0
            if len(seg) < 4:
                new_segs.append(seg + [0, 0, seg[0], seg[1]]) # why do you add these?
            if len(seg) == 4:
                new_segs.append(seg + [seg[0], seg[1]])
            else:
                new_segs.append(seg)
            new_segs.append(seg) # Why do we need this line?
        ann['segmentation'] = new_segs

Hi! Thank you very much for your excellent work!

How can I export the alpha mask of the segmented area? That is, a black and white image that only leaves the segmented area white?

Thank you very much in advance!

The blog post linked for demo page directs to detectron2 blog page - https://github.com/facebookresearch/VLPart/tree/main/demo

instead of - https://cheems-seminar.github.io/

Hello!

What weight are you saving for the zero shot weights?

Thank you for this great work! When I run your scripts to generate the annotations for pascal part I only get 8831 total samples as opposed to the expected ~10,000. Do you know why this would be?

Hi, I appreciate your great work, however, when I run evaluation on PACO dataset follow your instructions, I found out that both r_50 and swinbase model reach a higher AP than your paper on open-vocabulary task, the details are as follows:

r_50:

Swinbase:

and my evaluation command is:
python train_net.py --config-file configs/paco/r50_paco.yaml --eval-only --num-gpus 8
and:
python train_net.py --config-file configs/paco/swinbase_cascade_paco.yaml --eval-only --num-gpus 8

Could you please check your result or inform me that how should I reproduce your result, thanks!

Hi, I very appreciate your work and I would like to follow your work and apply it to recent tasks, but I wonder how could I visualize these image-level data that were parsed into part-level data, I found out that in configs folder there is another folder named ann_parser which contains the config file to parse data. However, when I use the config file "find_ins_mixer.yaml" it went wrong with "FileNotFoundError: [Errno 2] No such file or directory: 'output_basedata_pascalpart' ", I would like to know that what command should I run and how should I prepare datasets for visualizing those images that were parsed into part-level data, thanks.

First thank you for your great work!!

The model outputs several overlapped regions with a same label when a picture which doesn't belong to any dataset was input. I wonder why this happend and how to fix it?

Thank you in advance!

Hi @PeizeSun @ShoufaChen et al.,

Thx for ur great contribution. When I tried the demo on the image below, with the cmd:

python demo/demo.py --config-file configs/partimagenet_ablation/r50_lvis_paco_pascalpart_ins11_ins11parsed.yaml \
  --input input1.jpg \
  --output output_image \
  --vocabulary pascal_part \
  --confidence-threshold 0.7 \
  --opts MODEL.WEIGHTS models/r50_lvis_paco_pascalpart_ins11_ins11parsed.pth VIS.BOX False

I cannot get satisfactory generalization results: e.g., misclassify the leg as the arm. Could you provide some guidance on this issue? Many thx & rgds,

Refs:
https://github.com/facebookresearch/VLPart/blob/main/MODEL_ZOO.md#cross-dataset-part-segmentation-on-partimagenet

Thanks for the great work! I'm hoping to test VLPart for detecting objects from aerial images. However, The results are not as expected. Could you please help me? I wasn't sure if I was using the repo incorrectly or if this was the limitation of the released model on the aerial image. Thank you in advance for your help!

Here is a result I got by running :

python demo/demo.py --config-file configs/joint_in/swinbase_cascade_lvis_paco_pascalpart_partimagenet_inparsed.yaml --input 2.jpg --output output_image --vocabulary custom --custom_vocabulary "road,building,window,tree,car,light pole" --confidence-threshold 0.7 --opts MODEL.WEIGHTS models/swinbase_cascade_lvis_paco_pascalpart_partimagenet_inparsed.pth VIS.BOX False

I also tried to lower the confidence-threshold, but other than the cars, nothing was detected.