By Gideon Kowadlo and Abdelrahman Ahmed
The dataset is an integral part of an ML engineer’s toolkit. Like any tool, it’s important to choose your dataset carefully! One of the primary factors is how well known it is. The difficulties are well appreciated and performance benchmarks are available. You know what to expect, and others can evaluate your results.
Luckily, in Computer Vision these days, there are many publicly available datasets with published results. Comparing though, is surprisingly difficult. A lot of important details are often missing from the respective web pages and original papers, and require exploring of the dataset manually.
We recently compiled useful information about a range of these well known datasets. It’s all in one place, and hopefully useful to others as well. Features such as such as size, dimensions and perturbations are all in one place, making it easier to compare and choose the right dataset for an algorithm/experiment. Let us know if you find it useful, and if you have any suggestions for improvements.
Most of the columns are self explanatory. ‘Perturbations’ refer to the dataset as is without additional processing i.e. various instances of the same class at different rotations, occlusions, scale changes etc.