The First Place Solution of iMaterialist (Fashion) 2019

Solution

My solution is based on the COCO challenge 2018 winners article: https://arxiv.org/abs/1901.07518.

Model:

Hybrid Task Cascade with ResNeXt-101-64x4d-FPN backbone. This model has a metric Mask mAP = 43.9 on COCO dataset. This is SOTA for instance segmentation.

Validation:

For validation, I used 450 training samples splitted using https://github.com/trent-b/iterative-stratification.

Preprocessing:

I applied light augmentatios from the albumentations library to the original image. Then I use multi-scale training: in each iteration, the scale of short edge is randomly sampled from [600, 1200], and the scale of long edge is fixed as 1900.

Training details:

pre-train from COCO
optimizer: SGD(lr=0.03, momentum=0.9, weight_decay=0.0001)
batch_size: 16 = 2 images per gpu x 8 gpus Tesla V100
learning rate scheduler:

if iterations < 500:
   lr = warmup(warmup_ratio=1 / 3)
if epochs == 10:
   lr = lr ∗ 0.1
if epochs == 18:
   lr = lr ∗ 0.1
if epochs > 20:
   stop

training time: ~3 days.

Parameter tuning:

After the 12th epoch with the default parameters, the metric on LB was 0.21913. Next, I tuned postprocessing thresholds using validation data:

rcnn=dict(
    score_thr=0.5,
    nms=dict(type='nms', iou_thr=0.3),
    max_per_img=100,
    mask_thr_binary=0.45
)

This improved the metric on LB: 0.21913 -> 0.30011.

Test time augmentation:

I use 3 scales as well as horizontal flip at test time and ensemble the results. Testing scales are (1000, 1600), (1200, 1900), (1400, 2200).

I drew a TTA scheme for Mask R-CNN, which is implemented in mmdetection library. For Hybrid Task Cascade R-CNN, I rewrote this code.

This improved the metric on LB: 0.30011 -> 0.31074.

Ensemble:

I ensemble the 3 best checkpoints of my model. The ensemble scheme is similar to TTA.

This improved the metric on LB: 0.31074 -> 0.31626.

Attributes:

I didn't use attributes at all: they were difficult to predict and the removal of classes with attributes greatly improved the metric.

During the whole competition, I deleted classes with attributes: {0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12} U {27, 28, 33}. But two days before the end I read [the discussion] (https://www.kaggle.com/c/kaggle-imaterialist-fashion-2019-FGVC6/discussion/94811#latest548137) and added back classes {27, 28, 33 }.

This improved the metric on LB: 0.31626 -> 0.33511.

Postprocessing for masks

My post-processing algorithm for avoid intersections of masks of the same class:

def hard_overlaps_suppression(binary_mask, scores):
    not_overlap_mask = []
    for i in np.argsort(scores)[::-1]:
        current_mask = binary_mask[..., i].copy()
        for mask in not_overlap_mask:
            current_mask = np.bitwise_and(current_mask, np.invert(mask))
        not_overlap_mask.append(current_mask)
    return np.stack(not_overlap_mask, -1)

Small postprocessing:

I deleted objects with an area of less than 20 pixels.

This improved the metric on LB: 0.33511 -> 0.33621.

How to run?

Docker

make build
make run-[server-name]
make exec

Build mmdetection:

cd mmdetection
bash compile.sh
python setup.py develop

Prepare pretrained weights:

bash prepare_weights.sh

Data structure

/data/
├── train/
│   └── ...
├── test/
│   └── ...
└── train.csv.zip
/dumps/
└── htc_dconv_c3-c5_mstrain_x101_64x4d_fpn_20e_1200x1900/

Fix the error in train.csv.zip.

Prepare annotations for mmdetection:

cd scripts
bash create_mmdetection_train.sh
bash create_mmdetection_test.sh
bash split.sh

Training the model:

CUDA_VISIBLE_DEVICES=[list of gpus] bash dist_train.sh [config] [gpus] [--validate]

Test the model:

CUDA_VISIBLE_DEVICES=[list of gpus] bash dist_test_ensemble.sh [config] [gpus]

References

https://github.com/open-mmlab/mmdetection

ArtechStark / kaggle-imaterialist