sejong-rcv / mlpd-multi-label-pedestrian-detection Goto Github PK

As shown in the figure below, I have a memory explosion problem when running the code, there is no problem with the model training, but when running to val_epoch there will be memory explosion, that is, when running to the line of detect_object() in the inference.py there will be not enough memory.

Hi,
How can i test the following model on unpaired images , i can't found where do i need to make changes to make it work.
Thanks in advance.

Hi! Thanks for your great job!
I want to reproduce the performance of cvc data based on your code.
I found the data related to the cvc dataset in the issue, there are only annotations for testing(such as 'CVC-14_test_all_Full.json') but no annotations for training(CVC-14_train_all_Full.json).

Can you provide relevant json file?
Have a nice day！🙂

I am puzzled about the height range defined in evaluation_script.py, i.e., self.HtRng of class KAISTParams.
It seems to set the targets with height less than 50 to be ignored in evaluation process.
And if we set the lower bound of height range to 16 or 30, which is lower than the classic boundary 32*32 for small targets in COCO, it can be found an evident decay in detection precison, i.e., an evident increasing of MR.
Could you please give the reason for the specific setting of height range ?

Use the code provided by the author for training, which is obtained
MR_all: 98.65
MR_day: 99.52
MR_night: 96.37
Recall_all: 6.53
May I ask what does MR mean? Why is it that recall is getting lower and lower when I retrain? What do these mean respectively and why are they different from the data displayed in your paper
MR_all: 7.58
MR_day: 7.96
MR_night: 6.95
Recall_all: 96.70
Different, Is there something I didn't adjust，a little confused, looking forward to your answer, thank you

Since only pedestrians larger than 55 pixels are evaluated for MR metric following "Multispectral Pedestrian Detection: Benchmark Dataset and Baseline", I am wondering about the preprocessing details for your KAIST training set.
That is, in the training phase, only pedestrians larger than 55 pixels will be trained or pedestrians of all sizes will be trained?

Thank you for your previous tracking on all the issues.
Since your released codes mainly focus on KAIST, could you please provide with the necessary preprocessing details of CVC-14 dataset conducted by MLPD?

Besides, your paper points out that "However, the author of the dataset released the cropped image pairs without the non-overlapped areas. Therefore, we treated this dataset as a fully-overlapped (paired) dataset for our purpose,...".
Thus I am wondering if the CVC-14 for MLPD are cleaned by filtering out the targets not belong to "the cropped image pairs without the non-overlapped areas" for training?

Thank you for your work.

I have a error while running inference.py
It is
AttributeError: module 'torch' has no attribute 'permute'
So I changed code
original_image = original_image.permute(1, 2, 0).contiguous()
from
original_image = torch.permute(original_image.squeeze(), (1, 2, 0))

but exist error about number of dims don't match in permute

I checked original image shape (1,1,512,640)

Is there any error while running??

Looking forward to your reply,

thank you

Hello, could you please provide the data loading code for training the CVC-14 dataset? The CVC-14 dataset cannot be trained in the current code.thanks.

Hi,

Thank you for your great work.
I have two questions about the category setting of the model.

1. I have noticed that the categories (classes) in KAIST-RGBT dataset annotation are ['ignore', 'person', 'cyclist', 'people', 'person?'], but the number of categories in the config file is set to three. What exactly does each category id mean in the model?

2. If my understanding is correct, the evaluation script converts the predicted result into a json format for submission to the evaluation server, where all the predictions are made to have a category_id of 1. Are the results of the model's detection considered as only one category because all the classes predicted by the model are sub-categories of 'pedestrian'?

Thank you.

Hi , since i am very interested in your work, i like to reproduce it with my own dataset but i got this error .

Error : stack expects each tensor to be equal size, but got [3, 720, 1280] at entry 0 and [3, 512, 640] at entry 1
(NB : my image dimension is 1280x720)

Thank you for your reply, since we are very interested in your work, we hope to reproduce the data in the paper. Based on your previous suggestion, our issue remains unresolved. Below I describe the work we do in detail, and hope to get your suggestions. We first resized the images on the basis of the original dataset, unified to 640×471, and then we merged the contents of the FramesNeg and FramesPos folders in the training set into the Frames file. After that we trained using the datasets.py, CVC-14_test_all_Full.json, Images_train_all.txt and Images_test_all.txt you provided, since we still used the config.py file for the kaist dataset, we did the datasets.py file Modified as follows;

In addition, we also modified some details in the eval code;

Although it can be trained, it can't go down during the day as described in the previous question. In addition, according to your suggestion, we only use the bounding box in the gray image for training and testing, and no longer convert the image in visible to RGB during training. , but the effect is worse. Can you give further advice on our work? Thank you so much, looking forward to your reply!

W: GPG error: https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64 InRelease: The following signatures couldn't be verified because the public key is not available: NO_PUBKEY A4B469963BF863CC
E: The repository 'https://developer.download.nvidia.com/compute/cuda/repos/ubuntu1804/x86_64 InRelease' is not signed
how to fix it?

Dear, thank you very much for the dataset.py file provided, but we still can't train on the CVC dataset because of the lack of something like kaist_annotations_test20.json, test-all-20.txt, train-all-02.txt, config.py , could you please provide these documents again, thank you.

Hi ,
I would like to know more about the two types of unpaired cases: RGB blackout and Thermal blackout.
I have read the paper but on the code it seems like you test on both visible and thermal images even we mention the option --FDZ blackout_r or blackout_t when running the inference file.
On visualisation the visible or thermal image depending on the case are fill zero values but on evaluation you work with both images visible and thermal, so How the testing of inference working on in the case of unpaired images?

Thanks in advance.