kaggle-pku-autonomous-driving

Part of 5th place solution for Peking University/Baidu - Autonomous Driving on Kaggle (https://www.kaggle.com/c/pku-autonomous-driving).

Solution

My approach is based on CenterNet. I reimplemented CenterNet almost from scratch with reference to the author's implementation.

Heads

• heatmap[1]
• xy offset[2]
• z (depth)[1]
• pose[6]: cos(yaw), sin(yaw) cos(pitch), sin(pitch), cos(roll), sin(roll)
• wh[2]: It's not used for prediction, but PublicLB was improved by learning this as an auxiliary task.

Heatmap's loss is Focal Loss, and the others are L1Loss. The weight of wh loss is 0.05. Mask regions of mask images are ignored when calculating loss.

Network Architecture

• ResNet18 (pretrained ImageNet) + FPN (channels: 256->128->64)
• DLA34 (pretrained KITTI 3DOP) + FPN (channels: 256->256->256)
• Input size: 2560 x 2048 (2560 x 1024)
• Output size: 640 x 512 (640 x 256)

Increasing the input size is very effective, mAP was improved dramatically. I tried deeper networks (ResNet34, 50) but not worked.

Augmentation

• HFlip (p=0.5): Flip images horizontally and yaw *= -1, roll *= -1.
• RandomShift (p=0.5, limit=0.1): Shift images and positions (x, y).
• RandomScale (p=0.5, limit=0.1): Scale images and positions (x, y, z).
• RandomHueSaturationValue (p=0.5, hue_limit=20)
• RandomBrightness (p=0.5, limit=0.2)
• RandomContrast (p=0.5, limit=0.2)

Training

• Optimizer: RAdam
• LR scheduler: CosineAnnealingLR (lr=1e-3 -> 1e-5)
• 50epochs
• 5-folds cv
• Batch size: 4

Post Processing

• Remove mask regions from predictions by multiplying heatmap by masks.
• NMS (distance threshold: 0.1): I'm not sure how effective this is...
• Find duplicate images with imagehash and ensemble them. PublicLB was slighly improved.
• Score threshold: 0.3 (for val mAP: 0.1)

Ensemble

Ensemble each fold models and two models (ResNet18, DLA34) by averaging the raw output maps.

Score Summary

model	val mAP (tito's script)	PublicLB	PrivateLB
ResNet18 + FPN	0.257224305900695	0.118	0.109
DLA34 + FPN	0.2681900383192367	0.118	0.112
Ensemble	0.27363538075234406	0.121	0.115

What Didn't Work

• Pseudo labeling
• TTA (hflip)
• Weight Standardization
• Group Normalization
• Deformable Convolution V2
• Quaternion + L1Loss
• Very large input size (3360 x 2688)
• Eigen's depth prediction method used in CenterNet paper (z = 1 / sigmoid(output) − 1)

Run

python train.py --name resnet18_fpn --arch resnet18_fpn
python test.py --name resnet18_fpn
python train.py --name dla34_ddd_3dop --arch dla34_ddd_3dop --num_filters 256,256,256
python test.py --name dla34_ddd_3dop
python ensemble_test.py --models dla34_ddd_3dop,resnet18_fpn