Fix COCO configuration

[chienyiwang]

I tried the training and testing on COCO dataset and fixed some minor bugs in the configuration.

[d-li14]

Thanks for your modification. But I wonder that how you deal with COCO dataset? The url https://s3.amazonaws.com/amdegroot-datasets/instances_trainval35k.json.zip given in data/scripts/COCO2014.sh works well? Or is there any link to download correct json file for splitting the 35k? @chienyiwang

[ddonatien]

Hi @d-li14 I hope you have solved your problem by now, but just for the record you can find other sources of the file as stated in issue #205 and #123

[d-li14]

Thanks @ddonatien , but as far as I'm concerned, as stated in #123, just replacing instances_trainval_35k.json with valminusminival2014.json seems not reasonable.

[foralliance]

@d-li14
@ddonatien

replacing instances_trainval_35k.json with valminusminival2014.json is wrong.

Theoretically, the instances_trainval_35k.json should contain 118287 images. If you do this substitution, the instances_trainval_35k.json only contain 35185 images

[d-li14]

@foralliance Yeah, I agree with you, but is there any available source of the correct json file?

[foralliance]

@d-li14
refer RFBNet

[d-li14]

@foralliance Well, thanks!

[SalahAdDin]

It does not work to me then, the script is asking fot the json file, but it is a zip that cannot be unzipped.

[Damon2019]

I tried the training on COCO dataset ,can anyone tell me how to modify the coco_labels.txt?

[814247846]

I think :1.trainval35k is COCO 2017 train,because it have 118287 images.
2.So the instances_trainval35k.json is instances_train2017.json

[SalahAdDin]

@814247846 Did you get work with that json file?

[814247846]

@814247846 Did you get work with that json file?

yes.But you also have to change:
img = cv2.imread(osp.join(self.root, path))---->img = cv2.imread(path)
I think it's possible that the author made a mistake.

[SalahAdDin]

@814247846 all PR are welcomed.

[MiChuan]

The data sets include 2014 (2014 Train images, 2014 Val images, 2014 Test images), 2015 (2015 Test images) and 2017 (2017 Train images, 2017 Val images, 2017 Test images). Among them, training set and validation set There are tags (2014 Train/Val annotations, 2017 Train/Val annotations), while the test set has no tags.

In the 2014 data set, the training set is 82,783, the validation set is 40504, and the test set is 40775. In addition, the validation set is divided into two parts, miniVal has 5000, and the remaining 35504 images and training set are called Trainval35k (Trainval35k = = train2014 + val2014 - minival2014.). Trainval35k is usually used as the training set in the paper.

There are 81,434 images in the 2015 test set.

In the 2017 data set, the training set is 118287, the verification is 5000, and the test is 40670. Among them, the training set is Trainval35k (82783 + 40504 - 5000 = 118287), and the verification set is miniVal. That is, train2017 == trainval35k == train2014 + val2014 - minival2014 == train2014 + val2014 - val2017.

Original link: https://blog.csdn.net/qq_37643960/article/details/103466904