Can not show custom data correctly.

[feiyuhuahuo]

My mmdetection version is 2.10.0. Following this guidance(https://github.com/open-mmlab/mmdetection/blob/master/docs/2_new_data_model.md), I prepared my custom dataset as a coco-style json. But I found the code can not read my data correctly. Because the CLASSES is fixed,

mmdetection/mmdet/datasets/coco.py

Line 32 in af95704

CLASSES = ('person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus',

so adding a custom class variable in the config file is useless.

All my images were filtered because of the incorrect classes.
Besides, I did not found the class variable in the default config.
I finally fixed this error by writing a custom Dataset. I tried to show my data by:python tools/misc/browse_dataset.py [config], but the showed image was broken:

I guess this is because my image is PNG file. But PNG format is important. In industry environment, images are usually in this format because of its lossless property. I haven't fixed this yet. Hope you can help me, thanks.

Additionally, I have a personal suggestion. Maybe it's better to open this min_size in config file. Or in some small object detection task, target maybe filtered.

mmdetection/mmdet/datasets/coco.py

Line 98 in af95704

def _filter_imgs(self, min_size=32):

[hhaAndroid]

Hi @feiyuhuahuo, thank you for your feedback.

1. If you have a custom dataset as a coco-style JSON. Your configuration should look like this：

dataset_type = 'CocoDataset'
classes = ('balloon', )

2. The picture is not displayed correctly because the image data type is float32. We will fix this problem

Thank you!

[hhaAndroid]

About the min_size in the config file, It means that images smaller than this value will be filtered instead of the target.

[feiyuhuahuo]

Writing a classes variable in config is useless. self.CLASSES is still coco 80 classes. Take a look at this line.

mmdetection/mmdet/datasets/coco.py

Line 58 in af95704

self.cat_ids = self.coco.get_cat_ids(cat_names=self.CLASSES)

[hhaAndroid]

Writing a classes variable in config is useless. self.CLASSES is still coco 80 classes. Take a look at this line.

mmdetection/mmdet/datasets/coco.py

Line 58 in af95704

self.cat_ids = self.coco.get_cat_ids(cat_names=self.CLASSES)

Your understanding is incorrect. This parameter will be modified during the initialization in CustomDataset

[feiyuhuahuo]

mmdetection/tools/misc/browse_dataset.py

Line 67 in af95704

dataset = build_dataset(cfg.data.train)

I've taken a look at this line. The classes is in the whole config, but is missed in cfg.data.train. Did I make something wrong?

[hhaAndroid]

Can you provide the complete config file content?

[feiyuhuahuo]

The whole config.

_base_ = '../_base_/default_runtime.py'
# model settings
model = dict(
    type='YOLOV3',
    pretrained='open-mmlab://darknet53',
    backbone=dict(type='Darknet', depth=53, out_indices=(3, 4, 5)),
    neck=dict(
        type='YOLOV3Neck',
        num_scales=3,
        in_channels=[1024, 512, 256],
        out_channels=[512, 256, 128]),
    bbox_head=dict(
        type='YOLOV3Head',
        num_classes=4,
        in_channels=[512, 256, 128],
        out_channels=[1024, 512, 256],
        anchor_generator=dict(
            type='YOLOAnchorGenerator',
            base_sizes=[[(1000, 150), (1000, 40), (500, 50)],
                        [(100, 150), (150, 30), (30, 100)],
                        [(40, 15), (40, 40), (15, 25)]],
            strides=[32, 16, 8]),
        bbox_coder=dict(type='YOLOBBoxCoder'),
        featmap_strides=[32, 16, 8],
        loss_cls=dict(
            type='CrossEntropyLoss',
            use_sigmoid=True,
            loss_weight=1.0,
            reduction='sum'),
        loss_conf=dict(
            type='CrossEntropyLoss',
            use_sigmoid=True,
            loss_weight=1.0,
            reduction='sum'),
        loss_xy=dict(
            type='CrossEntropyLoss',
            use_sigmoid=True,
            loss_weight=2.0,
            reduction='sum'),
        loss_wh=dict(type='MSELoss', loss_weight=2.0, reduction='sum')),
    # training and testing settings
    train_cfg=dict(
        assigner=dict(
            type='GridAssigner',
            pos_iou_thr=0.3,
            neg_iou_thr=0.3,
            min_pos_iou=0)),
    test_cfg=dict(
        nms_pre=1000,
        min_bbox_size=0,
        score_thr=0.05,
        conf_thr=0.005,
        nms=dict(type='nms', iou_threshold=0.45),
        max_per_img=100))

# dataset settings
# dataset_type = 'Haiboguan'
dataset_type = 'CocoDataset'
classes = ('huashang', 'yashang', 'yiwu', 'yise')
data_root = '/home/feiyu/Data/haiboguan/coco_format/'
img_norm_cfg = dict(mean=[0, 0, 0], std=[255., 255., 255.], to_rgb=True)
train_pipeline = [
    dict(type='LoadImageFromFile', to_float32=True),
    dict(type='LoadAnnotations', with_bbox=True),
    # dict(type='PhotoMetricDistortion'),
    # dict(
    #     type='Expand',
    #     mean=img_norm_cfg['mean'],
    #     to_rgb=img_norm_cfg['to_rgb'],
    #     ratio_range=(1, 2)),
    # dict(
    #     type='MinIoURandomCrop',
    #     min_ious=(0.4, 0.5, 0.6, 0.7, 0.8, 0.9),
    #     min_crop_size=0.3),
    dict(type='Resize', img_scale=[(1216, 384)], keep_ratio=True),
    dict(type='RandomFlip', flip_ratio=0.5),
    dict(type='Normalize', **img_norm_cfg),
    dict(type='Pad', size_divisor=32),
    dict(type='DefaultFormatBundle'),
    dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels'])
]
test_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(
        type='MultiScaleFlipAug',
        img_scale=(1216, 384),
        flip=False,
        transforms=[
            dict(type='Resize', keep_ratio=True),
            dict(type='RandomFlip'),
            dict(type='Normalize', **img_norm_cfg),
            dict(type='Pad', size_divisor=32),
            dict(type='ImageToTensor', keys=['img']),
            dict(type='Collect', keys=['img'])
        ])
]
data = dict(
    samples_per_gpu=8,
    workers_per_gpu=4,
    train=dict(
        type=dataset_type,
        ann_file=data_root + 'annotations/haiboguan_train.json',
        img_prefix=data_root + 'train/',
        pipeline=train_pipeline),
    val=dict(
        type=dataset_type,
        ann_file=data_root + 'annotations/haiboguan_val.json',
        img_prefix=data_root + 'val/',
        pipeline=test_pipeline),
    test=dict(
        type=dataset_type,
        ann_file=data_root + 'annotations/haiboguan_val.json',
        img_prefix=data_root + 'val/',
        pipeline=test_pipeline))
# optimizer
optimizer = dict(type='SGD', lr=0.001, momentum=0.9, weight_decay=0.0005)
optimizer_config = dict(grad_clip=dict(max_norm=35, norm_type=2))
# learning policy
lr_config = dict(
    policy='step',
    warmup='linear',
    warmup_iters=1500,  # same as burn-in in darknet
    warmup_ratio=0.1,
    step=[40, 60])

# runtime settings
runner = dict(type='EpochBasedRunner', max_epochs=70)
evaluation = dict(interval=2, metric=['bbox'])

[hhaAndroid]

Hi, Your config settings are incorrect. The correct config is as follows

train=dict(
        type=dataset_type,
        classes=classes,  #==========
        ann_file=data_root + 'annotations/haiboguan_train.json',
        img_prefix=data_root + 'train/',
        pipeline=train_pipeline),
    val=dict(
        type=dataset_type,
        classes=classes, #==========
        ann_file=data_root + 'annotations/haiboguan_val.json',
        img_prefix=data_root + 'val/',
        pipeline=test_pipeline),
    test=dict(
        type=dataset_type,
        classes=classes, #==========
        ann_file=data_root + 'annotations/haiboguan_val.json',
        img_prefix=data_root + 'val/',
        pipeline=test_pipeline))

[feiyuhuahuo]

Thanks, it works.