Roman Solovyev

You can download it from here: https://storage.googleapis.com/openimages/v5/class-descriptions-boxable.csv

Warning is ok. You have zero files in folder. Check if there are some JPGs available.

Typical errors: 1) Check preprocessing of images 2) Check min/max size of images 3) Check you read it BGR Also reduce THR for predicted boxes I propose to check score...

As I can see from LB you already solved the problem?

You can install keras-retinanet from pip: `pip install keras-retinanet`

If you use generator from this repository, it already has some color augmentations on this line: https://github.com/ZFTurbo/Keras-RetinaNet-for-Open-Images-Challenge-2018/blob/master/retinanet_training_level_1/open_images_generator.py#L347

Yes. It's the exact augmentations I used previous year. It works fine. Morphological augmentation was used from original RetinaNet repository. This year I rewrote them to use [albumentation ](https://github.com/albu/albumentations)library.

Can you please add more details of what are you trying to do.

May be this question is more suitable for keras-retinanet repository: https://github.com/fizyr/keras-retinanet To train on different labels you can use pretrained model from this page. Check FAQ how to do so:...

I used this function. I found it somewhere on github, can't remember where exactly ``` def cpu_soft_nms(boxes, sigma=0.75, Nt=0.5, threshold=0.01, method=2): N = boxes.shape[0] pos = 0 maxscore = 0...