Detectron: train with cityscapes

[jamestodo]

I want to train a model with cityscapes by using the configs/getting_started/tutorial_1gpu_e2e_faster_rcnn_R-50-FPN.yaml.I've changed the
NUM_CLASSES as 9,and the cityscapes dataset in my disk look like the description in the lib/datasets/dataset_catalog.py.

And i converted the cityscapes's annotations format to coco annotations format by the lib/datasets/cityscapes/tools/convert_cityscapes_to_coco.py.

I trained the model by the cmd python tools/train_net.py --cfg ./configs/getting_started/tutorial_1gpu_e2e_faster_rcnn_R-50-FPN.yaml

After about 4 hours,the model trained done.And i got the output:
INFO test_engine.py: 212: Wrote detections to: /home/todo/Detectron/output/test/cityscapes_fine_instanceonly_seg_test/generalized_rcnn/detections.pkl INFO test_engine.py: 64: Total inference time: 172.142s INFO task_evaluation.py: 75: Evaluating detections WARNING task_evaluation.py: 83: Cityscapes bbox evaluated using COCO metrics/conversions INFO json_dataset_evaluator.py: 162: Writing bbox results json to: /home/todo/Detectron/output/test/cityscapes_fine_instanceonly_seg_test/generalized_rcnn/bbox_cityscapes_fine_instanceonly_seg_test_results.json INFO task_evaluation.py: 61: Evaluating bounding boxes is done! INFO task_evaluation.py: 180: copypaste: Dataset: cityscapes_fine_instanceonly_seg_test INFO task_evaluation.py: 182: copypaste: Task: box INFO task_evaluation.py: 185: copypaste: AP,AP50,AP75,APs,APm,APl INFO task_evaluation.py: 186: copypaste: -1.0000,-1.0000,-1.0000,-1.0000,-1.0000,-1.0000

The AP all equal to -1.0.Then i run the test demo.Here is the result:

I think it might something error that i didn't know and any suggestion is appreciated.
Thanks a lot!

[wuqi930907]

Can you tell me how to use 'lib/datasets/cityscapes/tools/convert_cityscapes_to_coco.py' to convert cityscape to coco format?
Thanks @jamestodo

[jamestodo]

FYI,according to the #111 .

[krumo]

Hi, did you solve this problem? I got the similar error. @jamestodo

Update:
I think the problem is because I was trying to compute AP on cityscapes_fine_instanceonly_seg_test dataset. Because Detectron will not do evaluation on any dataset whose name contains 'test', Detectron will return a default -1 for all AP. It will be fine if you evaluate on cityscapes validation dataset.

[na-hri]

Hi @jamestodo...were you able to solve your issue?

[lssily]

Hi! Where did you get the convert_cityscapes_to_coco.py?

[cy94]

@lssily It's here in case you are still interested

[ir413]

Closing as the issue seems resolved (see @krumo's updated comment above).