xmed-lab / clip_surgery Goto Github PK

As is in the title

As you know, tensor activations tend to act noisy on unseen input image. So for the clarity of segmentation, it would be better if you use masking based on tensor values. Please check on Pull requests

Thank you for sharing excellent work!

I am trying to get segmentation masks (open-vocabulary) from the code.
I tried argmax from "similarity_map" from demo.py, and it showed lower performance.

Is there any way to get a segmentation mask?

Hi! Thanks for your great work.
Could you please provide a detailed explanation of how the category weight 'w' in formula (7) is designed? Why is it crucial to emphasize obvious classes?

Thanks for your excellent work.

I failed to reproduce the multi-label recognition results in Table 7. For example, when I use CLIP ViT-B/16 with softmax function, I only got 35% mAP on NUS-Wide (42.85% in paper). I use the cls token of the original CLIP without feature surgery. Could you share the details and evaluation code of multi-label recognition?

I have an error in jupyter :

AttributeError: module 'clip' has no attribute 'encode_text_with_prompt_ensemble'

"encode_text_with_prompt_ensemble" is a function? How can i solve it ?

Thanks!

Hi, Thank you for great work !

I wonder if your method also can be applied for EVA-CLIP ! Have you ever tried?
Thanks to MIM training, EVA-CLIP-L showed 15 mIoU zero-shot performance on Cityscapes Validation, while original CLIP showed 0 mIoU without any scheme(i.e. MaskCLIP, CLIPSurgery).
Therefore, I wonder your Surgery method can also boost zero-shot EVA-CLIP too !

The segmentation effect is good when there are existed object in the image, but when a random prompts text is given, a high-confidence point will also be generated, which will cause segmentation errors in the downstream SAM model. How can this situation be solved?
here is some case：
I need to find the bag in the image , but the bag actually is not in the image,but the generated point is also high score.

Hi, thanks for your great work.

I am interested in the details about open-vocab segmentation and I have few questions regarding this task.

1. In the architecture surgery, I'm wondering whether the prediction for segmentation comes from the original path or the new path? Additionally, which features are used in the feature surgery? The paper said "Note that Eq. 9 is specifically designed for the explainability task", but I think the segmentation should use this too?

2. And it confused me in the [code](https://github.com/xmed-lab/CLIP_Surgery/blob/e346359d67e8fc4fe301467914151316d3982661/clip/clip_surgery_model.py#L349C36-L349C36)

   x[0, :, :] = x_ori[0, :, :] # clip_surgery
   

   Why do you preserve the [cls] token in the original_path? If my understanding was right, the [cls] token in the original_path is not influenced by the new_path. So for the multi-label recognition task, the architecture surgery would be useless?

3. Could you give more details? And it would be of great help if you could release the code for the open-vocabulary segmentation.

Thanks again for your work!

Hi there

Thanks for this good work!

I am trying to explore input a sentence to get some satisfied SAM results with CLIp surgery, could you please provide some suggestions on it?

Best.

I want to reimplement the eclip, can you give more details about it?

First of all great work and thank you for sharing!

I have a question that I am in the process of figuring out but I thought I would post it here incase it was useful. I was able to get everything working quite successfully but I ran into a case I had a problem with. As you can see below I have a photo with two birds, through clip.get_similarity_map I was able to select the two birds

After when I want to implement clip.similarity_map_to_points to infer for SAM, it only seems to provide me with one bird

As you can guess in the title, I think this is a limitation of similarity_map_to_points method, but I am not sure. Would you be able to provide some clarity on similarity_map_to_points method any/or maybe suggest why this is happening.

Again, great work and thank you for posting!

your work is very good.I'd like to know more details.Can you provide the paper?

Thanks for your excellent work, but I can't understand that the redundant features Fr can be obtained with equation 8. Can you help me with this question?

your work seems similar to Extract Free Dense Labels from CLIP

Hi, thanks for this great work, I've noticed that clip_surgery also achieves good performance in open-vocabulary segmentation. However, clip_surgery requires specific words as input to obtain the corresponding segmentation results. I'd like to ask how clip_surgery is implemented in the open-vocabulary segmentation setting?

Hi! thank you for this good work and neat implementation
Have you tried training/fine-tuning CLIP_surgery on out of domain datasets (medical scans, drawings ..etc)? Do you think that would improve the mIoU on these dataset or the model would collapse?

Thanks for your work and for sharing the code!

It is, however unclear to me how the mIoU was computed for the open-vocabulary segmentation tasks.
In the paper (sec.5.1.3 page 10), you mention that:

"Specifically, the mIoU is also used to measure the visualization quality. While each positive label is evaluated independently with a grid search threshold to identify the foreground."

I have several questions:

1. For each image, do you only evaluate the mIoU of the classes present in the image? i.e., if there is only a "cat" and "dog" in the image, you only compute the mIoU for those 2 classes?
2. It appears to me that you define one similarity threshold for every class. Hence, how do you find the optimal threshold?
3. Is the computed threshold the same across the dataset, or do you compute it for each image?

Thanks,

1. I run the demo with CLIP "CS-ViT-B/16", and the results of "person" is as following with wrong points:
   

Could you help me find my errors?

2. More, I try "a person on the bench" and other sentences. Does the results with a long sentence input or with some extra description not work?
3. Could you share the training dataset of "CS-ViT-B/16"?

Thanks a lot.

Hi, thanks for your great work! I just wonder have you tried CLIP Surgery on traditional classification datasets like ImageNet and Cifar?

Hello
Thanks for the great work.
Do you have a script to convert your CLIP_Surgery to onnx format ?

Great work, is that possible for this work extend to the medical imaging field?

Hi, Thank u for your great works! I just want to ask about the exact definition of $m_c$ and $m_s$ in metric mSC, could you provide a specific example to illustrate it? Thank you again!

你好，
论文中5.6不是很明白。clip在训练的时候是句子和图片pair。如何得到每个单词与图片的相似度的呢？等式(9)里的Nt，是text token数量，那么在5.6的实验中，把句子中的每个单词作为text token吗？

Hi, thanks for the awesome repo.

I had a question about how clip processes the images that are different then 224x224, specifically for the high resolution images in the demo.
When I load the vit-b-16 model, and print the shape for the model.visual.positional_embedding, it is 197x768. Then when I encode the high res image (512x512) using model.visual, I notice that the shape of the positional embeddings automatically change to greater then 197 to be compatible with the image tokens. Can you tell me where in the code the positional embeddings is changing dynamically with the input image?

Hi,

First of all, great work! I have implemented CLIP_Surgery in my project and can confirm that it's better than clipseg at certain tasks.

However, I'm having a hard time getting it to make decent selections of small objects when using SAM. Let me give you an example:

Source image:

CLIP_Surgery selection of "hand" without SAM:

CLIP_Surgery selection of "hand" with SAM (it selected the background?):

clipseg selection of "hand":

Now, SAM outputs 3 different masks but the one above was selected as having the highest score per masks[np.argmax(scores)]. But if I look at the outputs, I can see that it really should have preferred mask0 in this case:

Is this an issue with CLIP_Surgery's implementation of SAM or SAM itself?

Also, even the best SAM mask seems to include a lot of background noise not present in clipseg's output. Is there an easy way to filter that out?

Thanks!