fanq15 / few-shot-object-detection-dataset Goto Github PK

Hello,
Could you please elaborate on the process of validation? I couldn't get from the paper what exactly is the validation set for hyperparameter tuning?
Thanks

Thanks for your pioneering work on few-shot object detection. Would you mind giving descriptions with more details over the evaluation protocol, i.e. sampling method for test episodes with support & query images?

Hi, can you tell me how to get the 5-shot support set, depends on our split?

Hello, I am honored to see this open data set, thank you.I used your code today to report this error when generating the json file:

Traceback (most recent call last):
  File "train_converter.py", line 37, in
    assert os.path.exists(img_path)
AssertionError

I printed the key and value:

key=0,value={'father_category': 'cosmetics', 'image_path': './all_images/part_2/train_part_e/e8/e8fe15d38a65c26b.jpg', 'image_name': 'e8fe15d38a65c26b', 'dataset': 'openimage', 'width': 1024.0, 'bbox': [393, 347, 48, 42], 'height': 719.0, 'category_name': 'lipstick'}

I found that there is no such folder: part_2/train_part_e. May I ask whether the data set I downloaded is different from the one you used when writing the code?
The folder under my part2 is:

Hi@fanq15 Thank you for sharing the code! What's the version of torch and is ok in higher version? I meet some problem when import roi_pooling as follow:

Thank you! best regard!

Hi, what's the batch size when training your proposed detector ? Your default setting is 2-way-5-shot, so the input should be (5 support images + 1 query images + 5 negative support images) * 2. on each GPU card (the memory seems be an issue)? And the total iterations are 22*60k? I am trying to reimplement your work and try to query the details of your training settings. Can u show me ur details of training? Thx

Hi, in figure 5,there are 3 depth-wise conv,what are the kernel size of the first two depth-wise convs?Is also 1x1?Can u tell me? Thx.

For example, in fsod_train.json, the annotation with id = 903, it's bbox = [387.0, 595.0, 26.0, 39.0], the height and width of corresponding image with image_id = 2019000000903 are 533 and 800, obviously the bbox is out of range.

There are a few wrong bbox annotations like this.

Hello, fantastic job!
Two questions about evaluation.
(1) In the 5-shot detection scenario, you mix up all test images and for each object category, pick 5 images that contain the target object. Do you test all the remaining images to get a mAP or do you just take one image at a time as a query image and repeat it ( pick 5 images again and test another query image randomly ) multiple times?
(2) The same question applies to the 5-way 5-shot evaluation.

Hi@fanq15, I am interest in your job and when open the code? I have some problems as follow:
Computing the similarity about support and query feature map, in general, by the way that computing the the cosine distance then what's the adervantages about this paper way? what's the similar? a value or vector?

Would you mind giving some clarification about the difference between the code under lib/core and lib/core/test_engine? Mainly for "test.py" and "test_engine.py"? Thanks in advance.

Hi, thanks for your shared Few-Shot-Object-Detection-Dataset.
Could you share the source code of paper "Few-Shot Object Detection with Attention-RPN and Multi-Relation Detector"?

Waiting for your reply soon. :)

Hi Authors,

I am currently testing a similar setting on FSOD. May I know for the full 200 classes testing, did you iterate through all 200-class x 5 support images for each testing images? It seems to be very time consuming.

Thanks

Thank you very much for sharing your code. Now I have a question about fsod dataset. Firstly，From the data set of fsod, we can see that it is divided into Part1 and part2.Can you tell me the difference between the two？ Secondly，Where did we choose our query image? part1 or part2？Does the batch size we set refer to the number of query images per input?

Looking forward to your reply. Thank you！

Hi, after reading your paper, I'm confusing that how the RPN use the similarity map to build the proposal generation? In section 4.2.1 , the paper just says the proposals are generated based on the attention feature map. ( Is the same way as Siamrpn++ : use a 1*1 conv to adjust the response map channel to fit the regression and classification task separatedly ?)

Hi, should you provide such a mapping table?
E.g:
coco_id_name_map={1: 'person', 2: 'bicycle', 3: 'car', 4: 'motorcycle', 5: 'airplane',...}

hello, I have read your paper, it's amazing! Hope to see your code someday! Thanks~

Hi, fanq15:
Your paper "Few-Shot Object Detection with Attention-RPN and Multi-Relation Detector".
In Figure 4, before the RoI Pooling layer, how did the proposal boxe of support sample come from?
Waiting for your reply soon. :)

The directory structure you write in the README doesn't work. The actual structure should be:
YOUR_PATH
├── code
│ ├── training scripts
│ ├── testing scripts
│ ├── ...
│
└── fsod
├── all_images
│ ├── part_1
│ └── part_2
|
├── fsod_train_df.pkl
├── fsod_test_df.pkl
├── train_converter.py
└── test_converter.py

hello, I have read your paper, it's amazing! Hoping to see your code someday! Thanks~

You say ‘The proposal box of the support sample is from the GT label. Because we are given the support image and boxes to detect objects in the query images.’
In Fig 4 you mention that For N-shot training, we obtain all the support feature through the weight-shared network and use the average feature across all the supports belonging to the same category as its support feature.
But how can we use GT label in average feature？

1

Hey!

Thanks for all your efforts in putting up this dataset, it is definitely amazing! One small question: what's the order of bbox annotation? or is there an order (i.e. left, top, bottom, right in VOC) in bbox annotations?
I'm asking this question since I found that bbox ordering seems to be inconsistent, for example, the first three annotations in test are:
index 0
bbox: [120, 80, 43, 127]
width: 500
height: 499

index 1
bbox: [0, 0, 1022, 1022]
width: 1024
height: 1024

index 2
bbox: [219, 237, 43, 120]
width: 375
height: 500

Any advice would be very much appreciated, thanks!

A few follow up questions from the paper:

• During one-shot evaluation, did you test the configuration where the support image class was not present in the query image? I ask because in Figure 4 and from the text of the paper, it seems that the global head and local heads output match scores, however I didn't see an performance evaluation of how well the model can predict whether the support class was present in the query image.
• Table 3: I noticed that the performance with the attentional RPN was not included, what kind of performance did you observe in that configuration?

In Table 6 (Experimental results on FSOD test set for 200 novel categories with 5 supports evaluated in novel category detection.),

1. Did the first row mean Faster R-CNN pre-trained on ImageNet [56] and MS COCO [13], but not including FSOD? Then you fine-tune this model on the 200-way-5-shot support set?
2. Did the second row mean Faster R-CNN pre-trained on ImageNet [56] and MS COCO [13] and FSOD? Then you fine-tune this model on the 200-way-5-shot support set?
3. In the last row, what datasets did this model (your method) pre-trained on? what the difference between "trained directly" and the "❌ ✅" in the second column？

Thanks a lot for your reply!

When will you release the codes of experiments? I will appreciate your reply! Thanks.

Thanks for your great paper. I have a question about the three types of relation in Multi-Relation Detector.
I wonder if global realation means avgpooling the roi features and use cosine similarty to calculate the match score, but I don't know how to calculate the other two relations. Could please tell me the details about that?