Other datasets to predict bounding boxes on

I know that a good way to practice what we’ve learned is applying it to other datasets. I remembered that in the old part 1, Jeremy had done a notebook about the Kaggle fisheries competition, which someone labeled with bounding boxes to aid in classification (https://github.com/fastai/courses/blob/master/deeplearning1/nbs/lesson7.ipynb) Might be interesting to try to apply what we’re learning now to that same task!


The most popular one is COCO: http://cocodataset.org/#home


If you’re feeling up to it, there’s also ImageNet Object Localization Challenge which just opened up for submissions on Kaggle.

There’s also Kaggle’s Data Science Bowl. Although you’ll need to generate masks as the final output (something we haven’t covered as of yet).


I’m hoping Jeremy will touch on this at some point, as it seems to me like this (generating masks) involves classification of every pixel ( semantic segmentation perhaps ? )

I’m pretty sure we’re going to be covering segmentation, since there’s a notebook for the Carvana competition. Although from what I’ve read about the Data Science Bowl 2018, the problem is instance segmentation rather than semantic segmentation.

That’s true. I should be looking at other notebooks as well :joy:

lesson 8 on COCO 2017 dataset: https://github.com/tensoralex/misc_notebooks/blob/master/fastai_dl2_L8_coco_2017.ipynb


Nice @tensoralex! :slight_smile: You beat me to it :slight_smile: I am also planning on redoing lesson 8 on the coco 2017 dataset :slight_smile:


:slight_smile: i also wanted to try run it on resnext50, but couldn’t make it work yesterday night.

:+1: You mention the datasets links at the top. I’m assuming you only used the validation dataset zip for both validation and testing (as done in the lecture), or did you make use of the other downloads as well ?

yeah, i haven’t used “test2017.zip”

okies. i was thinking of giving it a go with the other pascal voc dataset. maybe this evening. :thinking:


btw, if I understood the online docs correctly, tqdm.monitor_interval = 0 disables the monitor thread.(Does this have a side effect of disabling the RuntimeError: Set changed size during iteration warning.

Seems to me like multiple threads are updating the tqdm counter for the warning to pop up so often(at least on my notebooks), but I’m not really sure.

that what i was trying to fix “RuntimeError: Set changed size during iteration warning.” by setting tqdm.monitor_interval = 0
Did not help. :frowning:

1 Like

Agreed. Doesn’t help.



OpenImages dataset is mentioned in YOLOv3 paper.


I try to use this one to detect cars and other object on the road: http://www.cvlibs.net/datasets/kitti/eval_object.php?obj_benchmark=2d


Lesson-8 applied to the tiny imagenet dataset here:

I had a few issues with bounding boxes and the geometry of the Linear layer in the new head, is working fairly well but needs a bit more qc.