open-images-2019-instance-segmentation

Codes for Open Images 2019 - Instance Segmentation competition using maskrcnn-benchmark.

The result is not outstanding but the solution might be valuable to be shared because it used the famous maskrcnn-benchmark library 'as it is' and also used its outputs as it is without TTA or any post processing.

The detailed solution can be found in kaggle discussion.

Preparation

1. Install maskrcnn_bencmark according to official guide.
2. Download the Open Images dataset to the project root directory (or make sim link).

PROJECT_ROOT
├── README.md
├── config
│   └── e2e_mask_rcnn_X_101_32x8d_FPN_1x.yaml
├── create_dataset.py
├── create_submission.py
├── datasets
│   ├── challenge-2019-label300-segmentable-hierarchy.json
│   ├── challenge-2019-train-segmentation-masks.csv
│   ├── test
│   ├── train
│   └── train_masks
├── test.py
├── train.py
└── util.py

• test: test images (99,999 images)
• train: train images
• train_masks: train mask images

Trained models are available from Kaggle Dataset. If you use the trained models, please skip to the 'Test for Layer 0 Classes' section.

Train on Open Images Dataset

Create Dataset for Layer 0 Classes

Create COCO format dataset for layer 0 class.

python create_dataset.py -l 0

The COCO format dataset is created as:

PROJECT_ROOT
├── datasets
│   └── coco
│       ├── annotations
│       └── train2017

This is the COCO-based format, thus it can be used on the other library like mmdetection (but not tested).

Train for Layer 0 Classes

Train on 8GPUs. This requires only 14 hours using V100 8GPUs

python -m torch.distributed.launch --nproc_per_node=8 train.py OUTPUT_DIR "layer0"

Train on a single GPU. This requires about 4 days using a V100 GPU.

python train.py --config-file config/e2e_mask_rcnn_X_101_32x8d_FPN_1x_1gpu.yaml OUTPUT_DIR "layer0"

Training steps can be reduced without large degradation of accuracy. The following should requires only a day for training with a single GPU.

python train.py --config-file config/e2e_mask_rcnn_X_101_32x8d_FPN_1x_1gpu.yaml OUTPUT_DIR "layer0" SOLVER.STEPS "(70000, 100000)" SOLVER.MAX_ITER 120000

Test for Layer 0 Classes

python test.py -l 0 --weight [TRAINED_WEIGHT_PATH (e.g. layer0/model_0060000.pth)]

The resulting files will created as:

PROJECT_ROOT
├── datasets
│   └── test_0_results

Create Submission File for Layer 0 Classes

python create_submission.py -l 0

Create Submission File for Layer 1 Classes

Do the same procedure also for layer 1 classes:

python create_dataset.py -l 1  # this overwrite layer 0 dataset. Please move it if needed later
python -m torch.distributed.launch --nproc_per_node=8 train.py OUTPUT_DIR "layer1"
python test.py -l 1 --weight [TRAINED_WEIGHT_PATH (e.g. layer1/model_0060000.pth)]
python create_submission.py -l 1

Integrate Two Submission Files

python integrate_results.py --input1 output_0.csv --input2 output_1.csv

OK, let's submit the resulting file integrated_result.csv !