I tried to upload a JSON file using COCO format with no luck,
Is there a document stating the upload format?
Posted by: tlong132 @ April 2, 2022, 9:53 p.m.Hi tlong132,
the scoring system expects a zip file with a single .json file (the results) in it. The results file should be formatted as described on the coco page: https://cocodataset.org/#format-data, with each prediction being represented as one annotation as defined in the "1. Object Detection" section. Additionally to the keys mentioned on the coco page, you have to add one "score"-key to each prediction. This score key should contain the confidence for the bounding box prediction in the range between 0 and 1.
For reference, you can take a look at the "dataset/results/results.json" file in the starting kit (participate -> files on the competition page).
We'll soon update the FAQs and add another link to a sample submission to make things clearer.
Cheers,
Mathias
Thanks for the reply,
I find that my uploaded results lead to Zero-score,
However when we check the qualitative results, It does show some good predictions,
I was wondering what are the file_id, and file_name correspondence of the validation files?
Is it possible to provide a prediction.json file with predictions['images'] and predictions['categories'] filled in and predictions['annotations'] leave blank?
Posted by: tlong132 @ April 6, 2022, 6:40 a.m.Hi tlong132,
I have checked your submission. Both, the categories array and the images array seem to be fine.
One thing that sometimes causes errors is the format of the bounding box coordinates: Our validation system expects [x,y, width, height], not [x1,y1,x2,y2] but your submission actually looks like you have been following the correct pattern.
However, we also updated the starting kit (participate -> files) to also contain a reference solution containing one dummy annotation in the correct format, the images array for the validation set, and the categories array. I hope this helps.
Best,
Mathias