taichi512

[eastchun]

Could you share taichi512.yaml and taichi512-metadata.csv if available?

[JialeTao]

Could you share taichi512.yaml and taichi512-metadata.csv if available?

Hi, I'm also intersted about this. Have you got the taichi512.yaml? And how did you prepare the TaiChi dataset with 512*512 resolution?

[eastchun]

No, I didn't get the taichi512 dataset. The reason why I closed this issue was I was no longer interested in Taichi dataset. BTW, in order to prepare taichi 512 dataset, please try the followings:

1) Refer to https://github.com/AliaksandrSiarohin/video-preprocessing

2) In load_videos.py,:

Comment out the following line:

    `save(os.path.join(args.out_folder, partition, path), entry['frames'], args.format)`

And, add the following:

        if len(entry['frames'])==0:  print('\nLength mismatch - video_id:', video_id)
        else: 
            # Save cropped video to out_folder
            save(os.path.join(args.out_folder, partition, path), entry['frames'], args.format)

Otherwise, it will raise error (and program exit) due to length mismatch for some videos.

3) Run load_videos.py by specifying

    --image_shape (512,512)        

Then, it will download all youtube taichi videos (as specified in taichi-metadata) and crop/store .mp4 videos (or .png images) (according to your specification) in 512x512 image size.

[JialeTao]

No, I didn't get the taichi512 dataset. The reason why I closed this issue was I was no longer interested in Taichi dataset. BTW, in order to prepare taichi 512 dataset, please try the followings:

1. Refer to https://github.com/AliaksandrSiarohin/video-preprocessing

2. In load_videos.py,: Comment out the following line:

   `save(os.path.join(args.out_folder, partition, path), entry['frames'], args.format)`

   And, add the following:

        if len(entry['frames'])==0:  print('\nLength mismatch - video_id:', video_id)
        else: 
            # Save cropped video to out_folder
            save(os.path.join(args.out_folder, partition, path), entry['frames'], args.format)

Otherwise, it will raise error (and program exit) due to length mismatch for some videos.

1. Run load_videos.py by specifying

   --image_shape (512,512)        

Then, it will download all youtube taichi videos (as specified in taichi-metadata) and crop/store .mp4 videos (or .png images) (according to your specification) in 512x512 image size.

Thanks, I have noticed this. And the main concern of mine is that, the raw taichi videos are cropped by the bbox parameters first to get human centered videos, and then resized to the given image shape. As seen In following figure. I found that some videos are of small sizes such as 300x300 after center cropping, and directly resize them to 512x512 may cause losses. The author said they remove some videos that lacks sufficient resolution to support this size. And I don't know the rules they filter videos.

Another question is that, almost all youtube links for downloading the dataset are broken, can you get the dataset?

[eastchun]

If the bbox is smaller than 512x512, black color is painted for those missing areas ... It will hurt the quality anyway. In the worst case, you just identify those cropped videos with black boundaries in the train/test folders and delete them.

BTW, I personally extracted and deleted those videos with broken links (as well as length mismatched) listed in taichi-metadata.csv. The new one is in the attachment.

On Sat, Jun 12, 2021 at 9:51 PM JialeTao @.***> wrote:

No, I didn't get the taichi512 dataset. The reason why I closed this issue was I was no longer interested in Taichi dataset. BTW, in order to prepare taichi 512 dataset, please try the followings:

1.

Refer to https://github.com/AliaksandrSiarohin/video-preprocessing 2.

In load_videos.py,: Comment out the following line:

`save(os.path.join(args.out_folder, partition, path), entry['frames'], args.format)`

And, add the following:

    if len(entry['frames'])==0:  print('\nLength mismatch - video_id:', video_id)
    else:
        # Save cropped video to out_folder
        save(os.path.join(args.out_folder, partition, path), entry['frames'], args.format)

Otherwise, it will raise error (and program exit) due to length mismatch for some videos.

1. Run load_videos.py by specifying

   --image_shape (512,512)

Then, it will download all youtube taichi videos (as specified in taichi-metadata) and crop/store .mp4 videos (or .png images) (according to your specification) in 512x512 image size.

Thanks, I have noticed this. And the main concern of mine is that, the raw taichi videos are cropped by the bbox parameters first to get human centered videos, and then resized to the given image shape. As seen In following figure. I found that some videos are of small sizes such as 300300 after center cropping, and directly resize them to 512512 may cause losses. The author said they remove some videos that lacks sufficient resolution to support this size. And I don't know the rules they filter videos.

[image: image] https://user-images.githubusercontent.com/38600167/121776484-b58c6080-cbbf-11eb-88a7-505bce006ddb.png

Another question is that, almost all youtube links for downloading the dataset are broken, can you get the dataset?

— You are receiving this because you modified the open/close state. Reply to this email directly, view it on GitHub https://github.com/snap-research/articulated-animation/issues/19#issuecomment-860049297, or unsubscribe https://github.com/notifications/unsubscribe-auth/AHACPT7I2AOUN5RSF3XIO4DTSNJ3JANCNFSM45KYVSVQ .

[JialeTao]

If the bbox is smaller than 512x512, black color is painted for those missing areas ... It will hurt the quality anyway. In the worst case, you just identify those cropped videos with black boundaries in the train/test folders and delete them. BTW, I personally extracted and deleted those videos with broken links (as well as length mismatched) listed in taichi-metadata.csv. The new one is in the attachment. … On Sat, Jun 12, 2021 at 9:51 PM JialeTao @.**> wrote: No, I didn't get the taichi512 dataset. The reason why I closed this issue was I was no longer interested in Taichi dataset. BTW, in order to prepare taichi 512 dataset, please try the followings: 1. Refer to https://github.com/AliaksandrSiarohin/video-preprocessing 2. In load_videos.py,: Comment out the following line: save(os.path.join(args.out_folder, partition, path), entry['frames'], args.format) And, add the following: if len(entry['frames'])==0: print('\nLength mismatch - video_id:', video_id) else: # Save cropped video to out_folder save(os.path.join(args.out_folder, partition, path), entry['frames'], args.format) Otherwise, it will raise error (and program exit) due to length mismatch for some videos. 1. Run load_videos.py by specifying --image_shape (512,512) Then, it will download all youtube taichi videos (as specified in taichi-metadata) and crop/store .mp4 videos (or .png images) (according to your specification) in 512x512 image size. Thanks, I have noticed this. And the main concern of mine is that, the raw taichi videos are cropped by the bbox parameters first to get human centered videos, and then resized to the given image shape. As seen In following figure. I found that some videos are of small sizes such as 300300 after center cropping, and directly resize them to 512*512 may cause losses. The author said they remove some videos that lacks sufficient resolution to support this size. And I don't know the rules they filter videos. [image: image] https://user-images.githubusercontent.com/38600167/121776484-b58c6080-cbbf-11eb-88a7-505bce006ddb.png Another question is that, almost all youtube links for downloading the dataset are broken, can you get the dataset? — You are receiving this because you modified the open/close state. Reply to this email directly, view it on GitHub <#19 (comment)>, or unsubscribe https://github.com/notifications/unsubscribe-auth/AHACPT7I2AOUN5RSF3XIO4DTSNJ3JANCNFSM45KYVSVQ .

Thanks, have you tested that how many tranining and testing videos are preserved after removing? And is it convenient for you to share me a copy of this dataset with 512 resolution?

[eastchun]

No, I didn't generate the taichi 512x512 dataset and I don't have it.

On Sat, Jun 12, 2021 at 11:16 PM JialeTao @.***> wrote:

If the bbox is smaller than 512x512, black color is painted for those missing areas ... It will hurt the quality anyway. In the worst case, you just identify those cropped videos with black boundaries in the train/test folders and delete them. BTW, I personally extracted and deleted those videos with broken links (as well as length mismatched) listed in taichi-metadata.csv. The new one is in the attachment. … <#m8847742225655814479> On Sat, Jun 12, 2021 at 9:51 PM JialeTao @.*> wrote: No, I didn't get the taichi512 dataset. The reason why I closed this issue was I was no longer interested in Taichi dataset. BTW, in order to prepare taichi 512 dataset, please try the followings: 1. Refer to https://github.com/AliaksandrSiarohin/video-preprocessing https://github.com/AliaksandrSiarohin/video-preprocessing 2. In load_videos.py,: Comment out the following line: save(os.path.join(args.out_folder, partition, path), entry['frames'], args.format) And, add the following: if len(entry['frames'])==0: print('\nLength mismatch - video_id:', video_id) else: # Save cropped video to out_folder save(os.path.join(args.out_folder, partition, path), entry['frames'], args.format) Otherwise, it will raise error (and program exit) due to length mismatch for some videos. 1. Run load_videos.py by specifying --image_shape (512,512) Then, it will download all youtube taichi videos (as specified in taichi-metadata) and crop/store .mp4 videos (or .png images) (according to your specification) in 512x512 image size. Thanks, I have noticed this. And the main concern of mine is that, the raw taichi videos are cropped by the bbox parameters first to get human centered videos, and then resized to the given image shape. As seen In following figure. I found that some videos are of small sizes such as 300300 after center cropping, and directly resize them to 512512 may cause losses. The author said they remove some videos that lacks sufficient resolution to support this size. And I don't know the rules they filter videos. [image: image] https://user-images.githubusercontent.com/38600167/121776484-b58c6080-cbbf-11eb-88a7-505bce006ddb.png Another question is that, almost all youtube links for downloading the dataset are broken, can you get the dataset? — You are receiving this because you modified the open/close state. Reply to this email directly, view it on GitHub <#19 (comment) https://github.com/snap-research/articulated-animation/issues/19#issuecomment-860049297>, or unsubscribe https://github.com/notifications/unsubscribe-auth/AHACPT7I2AOUN5RSF3XIO4DTSNJ3JANCNFSM45KYVSVQ .

Thanks, have you tested that how many tranining and testing videos are preserved after removing? And is it convenient for you to share me a copy of this dataset with 512 resolution?

— You are receiving this because you modified the open/close state. Reply to this email directly, view it on GitHub https://github.com/snap-research/articulated-animation/issues/19#issuecomment-860059485, or unsubscribe https://github.com/notifications/unsubscribe-auth/AHACPTYFZRL4CGD3PCDGS7TTSNT2BANCNFSM45KYVSVQ .

[JialeTao]

Thanks a lot. I'll have a try.