LVIS-for-mmdetection

support Large Vocabulary Instance Segmentation (LVIS) dataset for mmdetection

1. Install mmdetection

Follow the instruction of mmdetection to install

2. Install LVIS-API

To evaluate the LVIS, we also need its api. Follow the LVIS-API to install.

3. Prepare the LVIS Dataset

Download all the images and annotations from the LVIS homepage. Organize them in the data folder as:

mmdetection
├── mmdet
├── tools
├── configs
├── data
│   ├── LVIS
│   │   ├── lvis_v0.5_train.json
│   │   ├── lvis_v0.5_val.json
│   │   ├── images
│   │   │   ├── train2017
│   │   │   ├── val2017

4. Copy Code

• Copy LVIS.py and LVIS_utils.py into mmdet/datasets/ folder.
• Add from .LVIS import LVISDataset to mmdet/datasets/__init__.py, and 'LVISDataset' to __all__

5. Change the Config File

• You can refer to our htc_x101_32x4d_fpn_20e_16gpu_example.py.
• Change data, dataset_type and data_root.
• Set rcnn.score_thr=0.00001 and rcnn.max_per_img=300 in test_cfg.

6. A Weird logger.handlers Bug

When I was training the model, I found a weird bug that the logger won't print out anymore. It takes me a long time to figure out that it's caused by the logger.HasHandlers() function, which always returns the True even the logger.handlers = []. If you have the same problem, you can use logger.py to replace the original mmdet/utils/logger.py in mmdetection project. If you don't have this problem, never mind.