About training on Mapillary dataset.

[HCShi]

Could you please provide the training config files of Mapillary dataset?

[yanwei-li]

Hi, the difference between Mapillary Visita and COCO in config is given below:

MODEL:
 TENSOR_DIM: 400
 POSITION_HEAD:
    THING:
      POS_NUM: 3
      NUM_CLASSES: 37
      THRES: 0.2
      TOP_NUM: 200
    STUFF:
      NUM_CLASSES: 29
      THRES: 0.1
  SEM_SEG_HEAD:
    NUM_CLASSES: 29

  KERNEL_HEAD:
    INSTANCE_SCALES: ((1, 128), (64, 256), (128, 512), (256, 1024), (512, 4096),)
    TEST_SCALES: ((1, 128), (128, 256), (256, 512), (512, 1024), (1024, 4096),),
  INFERENCE:
    INST_THRES: 0.5
    SIMILAR_THRES: 0.95
DATASETS:
  TRAIN: ("mapillary_train_panoptic_separated",) # generate before
  TEST: ("mapillary_val_panoptic_separated",) # generate before
SOLVER:
  BASE_LR: 0.02
  CLIP_GRADIENTS:
    CLIP_VALUE: 15.0
  IMS_PER_BATCH: 32
  MAX_ITER: 150000
INPUT:
  MIN_SIZE_TRAIN: (1024, 1280, 1408, 1536, 1664, 1792, 1920, 2048)
  MIN_SIZE_TEST: 2048
  MAX_SIZE_TRAIN: 4096
  MAX_SIZE_TEST: 2048
  CROP:
    ENABLED: True
    TYPE: "absolute" # use with instance crop, which include instance in each random crop
    SIZE: (1024, 1024)

[HCShi]

Hi, the difference between Mapillary Visita and COCO in config is given below:

MODEL:
 TENSOR_DIM: 400
 POSITION_HEAD:
    THING:
      POS_NUM: 3
      NUM_CLASSES: 37
      THRES: 0.2
      TOP_NUM: 200
    STUFF:
      NUM_CLASSES: 29
      THRES: 0.1
  SEM_SEG_HEAD:
    NUM_CLASSES: 29

  KERNEL_HEAD:
    INSTANCE_SCALES: ((1, 128), (64, 256), (128, 512), (256, 1024), (512, 4096),)
    TEST_SCALES: ((1, 128), (128, 256), (256, 512), (512, 1024), (1024, 4096),),
  INFERENCE:
    INST_THRES: 0.5
    SIMILAR_THRES: 0.95
DATASETS:
  TRAIN: ("mapillary_train_panoptic_separated",) # generate before
  TEST: ("mapillary_val_panoptic_separated",) # generate before
SOLVER:
  BASE_LR: 0.02
  CLIP_GRADIENTS:
    CLIP_VALUE: 15.0
  IMS_PER_BATCH: 32
  MAX_ITER: 150000
INPUT:
  MIN_SIZE_TRAIN: (1024, 1280, 1408, 1536, 1664, 1792, 1920, 2048)
  MIN_SIZE_TEST: 2048
  MAX_SIZE_TRAIN: 4096
  MAX_SIZE_TEST: 2048
  CROP:
    ENABLED: True
    TYPE: "absolute" # use with instance crop, which include instance in each random crop
    SIZE: (1024, 1024)

Thanks for the reply! If I want to train on this dataset, should I prepare the dataset structure like follows?

mapillary_vistas/
training/
    images/
    instances/
    labels/
    panoptic/
  validation/
    images/
    instances/
    labels/
    panoptic/

Whether it is convenient for you to upload the corresponding files on training Mapillary dataset, thanks!

[yanwei-li]

Yes, the dataset structure is right. And the whole dataset is too large. Maybe you can refer to official data generation script and use the COCO-style to prepare the dataset.

[HCShi]

Yes, the dataset structure is right. And the whole dataset is too large. Maybe you can refer to official data generation script and use the COCO-style to prepare the dataset.

Thanks a lot! I'll try the scripts.

[HCShi]

Yes, the dataset structure is right. And the whole dataset is too large. Maybe you can refer to official data generation script and use the COCO-style to prepare the dataset.

Hi, I have generated the dataset structure followed the official, need I follow the command python datasets/prepare_panoptic_fpn.py, to extract semantic annotations from panoptic annotations? It seems that this dataset is not compatible with COCO.

[yanwei-li]

Hi, how to optimize 'Thing' and 'Stuff' is up to you. In prepare_panoptic_fpn.py, they view all things as a super-category for semantic segmentation. In our approach, we follow that in prepare_panoptic_fpn.py. The COCO-style JSON is suggested to be generated for prepare_panoptic_fpn.py by yourself.

[HCShi]

Hi, how to optimize 'Thing' and 'Stuff' is up to you. In prepare_panoptic_fpn.py, they view all things as a super-category for semantic segmentation. In our approach, we follow that in prepare_panoptic_fpn.py. The COCO-style JSON is suggested to be generated for prepare_panoptic_fpn.py by yourself.

Thanks a lot!