Skip to content

Latest commit

 

History

History
71 lines (50 loc) · 2.3 KB

dataset_setup.md

File metadata and controls

71 lines (50 loc) · 2.3 KB

Setting Up Datasets

Note that all datasets are downloaded on the TDT4265 compute server and the tulipan/cybele computers (saved under /work/datasets).

Index

Pascal VOC

For Pascal VOC dataset, make the folder structure like this:

VOCdevkit
|__ VOC2007
    |_ JPEGImages
    |_ Annotations
    |_ ImageSets
    |_ SegmentationClass
|__ VOC2012
    |_ JPEGImages
    |_ Annotations
    |_ ImageSets
    |_ SegmentationClass
|__ ...

You can download this if you are on the school network directly as a .zip file (Note that this dataset should be only used for educational/academic purposes).

With scp:

scp [ntnu-username]@oppdal.idi.ntnu.no:/work/datasets/VOC.zip .

unzip VOC.zip

Or you can download it from the PASCAL VOC website: http://host.robots.ox.ac.uk/pascal/VOC/

Note that we are using the VOC2007 TRAIN/VAL + VOC2012 TRAIN/VAL as the train set.

We use VOC2007 Test as the validation set.

TDT4265 Dataset

To take a peek at the dataset, take a look at visualize_dataset.ipynb.

Getting started

Download/Setup:

python3 scripts/update_tdt4265_dataset.py

This will automatically download the dataset on local computers and create symlinks to the folder data if you are on TDT4265 compute servers.

Using the larger dataset for task 2.5

You can use the new dataset by changing the dataset folder. No other change should be required. The config file ../configs/tdt4265_updated.py shows how to perform this change.

If you are using your own computer, you need to download the updated dataset. The updated download script (../scripts/update_tdt4265_dataset.py) will automatically download the dataset to your computer and save it to the folder: data/tdt4265_2022_updated.

If you want to download the old dataset, you can run:

python3 scripts/update_tdt4265_dataset.py --old

Dataset Information

Label format follows the standard COCO format (see here for more info).