upload the training and testing dataset

[zeruniverse]

To compare different methods, we must use the same training and testing dataset. As discussed, testing dataset is Microsoft coco. Should we download some images elsewhere as testing data for evaluation purpose?

I can create a release and collect & upload all training/testing dataset if you haven't done it @Lyken17

[Armour]

Is there any requirement for the training and testing images? (Like portrait or landscape or blahblah)

[zeruniverse]

Not really.

On Saturday, 5 November 2016, Chong Guo notifications@github.com wrote:

Is there any requirement for the training and testing images? (Like
portrait or landscape or blahblah)

—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub
#2 (comment),
or mute the thread
https://github.com/notifications/unsubscribe-auth/AEbvNJcd-1kNojEzsSL7HgjT-qWFqS9gks5q7XTggaJpZM4Kqe0t
.

[Lyken17]

I think you mistook my meaning. What I do is to split COCO dataset into Train / Valid set, rather using whole COCO dataset for testing.

We can take some life case pictures for demo/posters, but I don't think it is necessary to make a "new" dataset.

[zeruniverse]

You mean doing cross-validation? I think that's too slow. Why not just use
the whole coco as training set and find some other images as testing set.
Our task is not what coco designed for so I think it has no problem if we
use coco as training set and images elsewhere as testing set

On Saturday, 5 November 2016, Lyken Syu notifications@github.com wrote:

I think you mistook my meaning. What I do is to split COCO dataset into
Train / Valid set, rather using whole COCO dataset for testing.

We can take some life case pictures for demo/posters, but I don't think it
is necessary to make a "new" dataset.

—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub
#2 (comment),
or mute the thread
https://github.com/notifications/unsubscribe-auth/AEbvNCY-SNI18siS4X3HqeUh2LDsEa2jks5q7XZWgaJpZM4Kqe0t
.

[Lyken17]

No cross-validation.

For example, if COCO dataset has 6000 images. Then I pick 4000 as training set, 1000 for validation and 1000 for testing . And when switching to other method(s), the same 4000, 1000 and 1000 images are used for training, validation and testing respectively.

[zeruniverse]

It still means we should make it clear which 4000 are training images. Did
you already seperate them?

On Saturday, 5 November 2016, Lyken Syu notifications@github.com wrote:

No cross-validation.

For example, if COCO dataset has 6000 images. Then I pick 4000 as training
set, 1000 for validation and 1000 for testing . And when switching to other
method(s), the same 4000, 1000 and 1000 images are used for training,
validation and testing respectively.

—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub
#2 (comment),
or mute the thread
https://github.com/notifications/unsubscribe-auth/AEbvNLAiJqFAdKbUqGLOTbwn2G1qy0Sfks5q7XgegaJpZM4Kqe0t
.