Skip to content

Latest commit

 

History

History
91 lines (65 loc) · 3.44 KB

domain_adaption.md

File metadata and controls

91 lines (65 loc) · 3.44 KB
Raw

DAOD

In this part, we give the tutorial about domain adaptive object detection (DAOD).

Dataset

Tasks C2F C2B K2C S2C
Source(Labeled) Cityscapes Cityscapes KITTI Sim10k
Target(Unlabeled) Foggy-Cityscapes BDD100k-Daytime Cityscapes Cityscapes

Usage

Training

1. Use labeled data to train a baseline (aka "source only" model)

Before training，please download the pretrained backbone (vgg) to pretrained_model/backbone.

# |---------------------|--------|---------|
# | xonsh train_gpu2.sh | config | dataset |
# |---------------------|--------|---------|
# there are three dataset for DAOD baseline: C2F, K2C, S2C
# Note that C2B share the same "source only" model with C2F
cd examples/train/xonsh
xonsh train_gpu2.sh ./configs/baseline/baseline_uda.py C2F

  • In our implementation, we use 2-gpus to train.

  • You can also run bash train_baseline_uda.sh in examples/train/bash

  • After training, we organize the pretrained baseline to pretrained_model/baseline as follows:

    pretrained_model/
	└── baseline/
        ├── C2F.pth
        ├── K2C.pth
        └── S2C.pth

2. Use labeled data + unlabeled data to train detector

## there are four adaptation tasks: C2F, C2B, K2C, S2C
## C2F and C2B share the same "source only" model
cd examples/train/xonsh
xonsh train_gpu8.sh ./configs/labelmatch/labelmatch_uda.py C2F
  • In our implementation, we use 8-gpus to train.
  • You can also run bash train_uda.sh in examples/train/bash

Evaluation

# change "data_name" and "checkpoint" in scripts to support different dataset and trained model
cd examples/eval
xonsh eval_uda.sh

Performance

  • $\dagger$ is an ideal setting, using the label distribution from unlabeled data annotations.
  • mAP: AP50

Normal-to-foggy weather adaptation

C2F mAP truck car rider person train motor bicycle bus
source only 30.9 19.2 47.9 40.8 34.8 7.8 24.2 36.0 36.4
LabelMatch 52.4 42.0 62.2 55.4 45.3 55.1 43.5 51.5 64.1

Small-to-large scale dataset adaptation

C2B mAP truck car rider person train motor bicycle bus
source only 28.7 18.3 50.0 33.3 35.8 / 18.4 27.6 17.0
LabelMatch 38.8 39.4 54.6 37.4 42.9 / 25.7 29.8 41.7
LabelMatch$^\dagger$ 44.5 39.8 55.4 44.5 44.8 / 38.6 41.5 47.1

Cross-Camera adaptation & Synthetic-to-Real adaptation

K2C AP S2C AP
source only 42.2 source only 36.5
LabelMatch 51.0 LabelMatch 52.7
LabelMatch$^{\dagger}$ 52.2 LabelMatch$^{\dagger}$ 53.8