Deepdrive Dataset to TFRecord

Convert the Berkeley Deepdrive dataset to a TFRecord file. (Specifically for the Object Detection Task)

This repository shall help to create a tfrecord file for the berkeley deep drive dataset. I have no affiliation with Berkeley and/or the deep drive team.

Now also supports the new data format

Download dataset

1. Register at http://bdd-data.berkeley.edu/login.html . NOTE: The server does not serve an SSL ceritificate.

2. Go to the Download page http://bdd-data.berkeley.edu/portal.html#download

3. Accept the License and Download: "Images", "Labels"

4. Create a folder ~/.deepdrive/download

5. Place both zip files in that folder

Create dataset

You can use the script create_tfrecord.py in order to create the TFRecord file you need.

--fold_type = ['train', 'val', 'test'] : Select for which fold you want to create the tfrecord (default=train)

--version = ['100k', '10k'] : The Berkeley Deepdrive Dataset comes in two sizes. (default=100k)

--elements_per_tfrecord = integer : You can specify, how many images are put into one tfrecord file. Multiple TFRecord files are generated.

--number_images_to_write = integer : Restricts the number of files to be written. [E.g. to create smaller files to test overfitting]

--weather = str : Specify the weather which should be written to the tfrecord

--scene_type = str : Specify the scene_type which should be written to the tfrecord

--daytime = str : Restrict the daytime which should be written to the tfrecord

The resulting TFRecord files can be found in : ~/.deepdrive/tfrecord/[version]/[fold_type]/

Read dataset

Using read_data.py you can check your TFRecord file.

--batch_size = int: Specify the batch-size

--fold_type = see above

--version = see above

It will plot all images, and all boundingboxes.