Questions about the preprocessed dataset and the data used in the paper about objectron

[ahmadyan]

Hi Yunzhi, Good questions!

1. There are two data folders for the tf.records in the storage bucket: 1) objectron/v1/records and 2) objectron/v1/records_shuffled, in records, we extract every frame in the dataset (totallying 4M) frames, and in the records_shuffled, we just shuffle those samples. So if you want to randomly grab N frames from the dataset read the records_shuffled and .take(N).

2. For MobilePose, we re-trained the model on the dataset. The train/test split is available here. The original MobilePose paper is using a internal version of objectron (which also includes synthetic images, and the dataset size is bigger than Objectron). To provide a fair comparison, we re-trained MobilePose models on the public version and reported the numbers in the objectron papers (note the 3D IoU is slightly lower than the eval reported in MobilePose paper).

3. If you grab 1K image from the shuffled dataset, it will uniformly sample the dataset. That is good enough. In the shuffled dataset, there are different shards for training and test. e.g. chair_train_00001_of_0xxxx and chair_test_00001_of_0xxxx
   See records here: https://pantheon.corp.google.com/storage/browser/objectron/v1/records/chair

If you decide to regenerate the records from raw data, just make sure that you follow the same train/test split as ttps://github.com/google-research-datasets/Objectron/tree/master/index

[Uio96]

Great. Thanks for the detailed answer.