Garment Segmentation

Requirements

• pycocotools
• mxnet
• gluoncv

Note: please install nightly build gluoncv for less bugs

pip install gluoncv --pre --upgrade

Dataset

iMaterialist (Fashion) 2019 at FGVC6

• Convert imaterialist to COCO style

python tools/convert_imaterialist_to_coco.py --data_root datasets/imaterialist

• files structure

datasets
└─imaterialist
    ├── annotations
    │   ├── resize_rle_instances_val.json
    │   ├── rle_instances_train.json
    |   └── rle_instances_val.json
    ├── test
    ├── train
    ├── val
    ├── label_descriptions.json
    ├── sample_submission.csv
    └── train.csv

Training

Note: all using COCO pretrained models, only the categories less than 13 (main apparels) for segmentation.

• Mask R-CNN

python train_mask_rcnn.py --save-prefix train_logs/xxx/ --val-interval 1

• DeepLabV3

python train_segmentation.py --model deeplabv3 --dataset imaterialist \
                            --model-zoo deeplab_resnet101_coco --aux \
                            --checkname res101 --epochs 30 --lr 0.001 --ngpus 1 \
                            --workers 2 --batch-size 2 --test-batch-size 2

• DeepLabV3+

python train_segmentation.py --model deeplabv3plus --dataset imaterialist \
                             --model-zoo deeplab_plus_xception_coco --aux \
                             --base-size 576 --crop-size 512 \
                             --checkname xception --epochs 30 --lr 0.001 --ngpus 1 \
                             --workers 1 --batch-size 1 --test-batch-size 1 

Reference

• Train Mask RCNN end-to-end on MS COCO
• Reproducing SoTA on Pascal VOC Dataset