The shuffle operation in the data loading process. about df-net

[Yuliang-Zou]

Hi @wanghao14 , actually the consecutive frames are concatenated together during data preprocessing. So we can make sure we can get consecutive frames even if the data is shuffled.

[wanghao14]

@Yuliang-Zou Sorry, I still have some questions.
In the function named format_file_list, the path to the images is stored in the list file_list['image_file_list'], which looks like

['$data_root$/2011_09_26_drive_0001_sync_02/0000000002.jpg', '$data_root$/2011_09_26_drive_0001_sync_02/0000000004.jpg', '$data_root$/2011_09_26_drive_0001_sync_02/0000000005.jpg',..., '$data_root$/2011_10_03_drive_0042_sync_03/0000001166.jpg']

.
Then, it will be used as the input to tf.train.string_input_producer with shuffling, and I think the sequence looks like

['$data_root$/2011_09_26_drive_0001_sync_02/0000000005.jpg', '$data_root$/2011_10_03_drive_0042_sync_03/0000001066.jpg', '$data_root$/2011_09_30_drive_0028_sync_03/0000002128.jpg', '$data_root$/2011_09_26_drive_0001_sync_02/0000000036.jpg', ...]

, which means that the order of image sequence is not consecutive. Then in the function named unpack_image_sequence, the middle(3th) image is treated as target image and the first two images and last two images are treated as source images. I don't think these five images are consecutive, and this is where I am confused. Could you please explain this part? Or some details of your data preprocessing? Because I am new to tensorflow, my question may be naive. I hope it doesn't bother you.

[Yuliang-Zou]

You can see the data preprocessing in data/kitti/kitti_raw_loader.py. We actually concatenate consecutive frames before we start training.