Question about using my own data

[dradnats-kim]

hello,

I tried to use my own data (some images, and humandetection.json)

i encountered "No such file pr directory: ..... " error.

I read other issues, this error mostly related directory or config setting.

but I just replaced images in 'test_2017' and det file that I made.

//changed//

1. contents of "test_2017/". there are 15 images and their name are integers 1 to 15

2. "humandetection.json". tried to make similar with offered det file.
   score : (float), category_id : 1(int), bbox : [x,y,w,h] (list, float), image_id : (int)

//cause I guess//

1. image name or id

2. format of det file

3. something i didn't change

I look forward to hearing from you.

thank you.

p.s : when i use offered files and data, they work well

[GK4444]

how did u create ur own humandetection.json?

[dradnats-kim]

Hello, thanks for your reply.

this is my code making dic and human_detection.json

box = boxes[i]
cv2.rectangle(img,(box[1],box[0]),(box[3],box[2]),(255,0,0),2)
x, y, w, h = float(box[1]), float(box[0]), float(box[3]-box[1]), float(box[2]-box[0])

result = dict(score=(scores[i]), category_id=1, bbox=[x, y, w, h], image_id=int(img_id))
dump_results.append(result)

and I attached json file in txt format.
human_detection .txt

I tried to follow COCO format.

best regards,

[GK4444]

did u change the testset variable in config.py?

[dradnats-kim]

did u change the testset variable in config.py?

yes. I set COCO, testset in config.py like this

    ## dataset
    dataset = 'COCO' # 'COCO', 'PoseTrack', 'MPII'
    testset = test' # train, test, val (there is no validation set for MPII)

and locate images and det file at right dir in my thinking
(images are in 'coco/images/test2017/' and det file in '/coco/dets/')

[GK4444]

did u also trained on ur custom data or are u just trying to test the pretrained model on ur data?

[dradnats-kim]

I just tried to test the pretrained model on my data. (model is trained on COCO dataset)

I thought 'If I make data with same format, it will works' .... Is it wrong??

[dradnats-kim]

Oh, I understand that I need annot_id for load images.

If I change code to load images directly, Will it be Okay?
(don't use path from COCO annot)

or Will it make another problem?

[GK4444]

I haven't gone through the whole codebase, so i don't know if direct loading the images will work or not. Another issue might be the structure of the human_detection.json, but again i am not sure.

[dradnats-kim]

I haven't gone through the whole codebase, so i don't know if direct loading the images will work or not. Another issue might be the structure of the human_detection.json, but again i am not sure.

or, Do you have any idea for use custom dataset?

For now, i'll try it.

[GK4444]

I trained a model on my cutom dataset but didn't try testing, because i have no idea how to create the human_detection.json file for my custom dataset.
Do let me know if u solve ur issue.

[mks0601]

Run detectron2 on your data.

[GK4444]

Ok Thanks, I'll try that.

[dradnats-kim]

Run detectron2 on your data.

Thanks for your reply.

I have additional question.

When you use Posefix, Is pose estimation data from Detectron2 ?

[mks0601]

As written in the paper, any pose results can be used for the input. Please read the paper for more details.

[dradnats-kim]

As written in the paper, any pose results can be used for the input. Please read the paper for more details.

Understandably, I know that.. one of the advantages is I can use any estimation method using coco format.

Just because my attempts to use custom data(coco format) was failed, I wondered you used it directly.

I think it's lack of my ability. thank you.

[mks0601]

Sorry, I can't get what you meant by 'directly', but the output of detectron2 can be used for the input pose.
Don't blame yourself :)

[dradnats-kim]

Sorry, I can't get what you meant by 'directly', but the output of detectron2 can be used for the input pose.
Don't blame yourself :)

Thanks for your kindness. I'll try them