[ECCV 2022] Domain Adaptive Video Segmentation via Temporal Pseudo Supervision

[Paper] [Video Demo]

Highlights

TPS is 3x faster than previous DA-VSN in training while achieves SOTA in domain adaptive video segmentation task.

Abstract

Video semantic segmentation has achieved great progress under the supervision of large amounts of labelled training data. However, domain adaptive video segmentation, which can mitigate data labelling constraint by adapting from a labelled source domain toward an unlabelled target domain, is largely neglected. We design temporal pseudo supervision (TPS), a simple and effective method that explores the idea of consistency training for learning effective representations from unlabelled target videos. Unlike traditional consistency training that builds consistency in spatial space, we explore consistency training in spatiotemporal space by enforcing model consistency across augmented video frames which helps learn from more diverse target data. Specifically, we design cross-frame pseudo labelling to provide pseudo supervision from previous video frames while learning from the augmented current video frames. The cross-frame pseudo labelling encourages the network to produce high-certainty predictions which facilitates consistency training with cross-frame augmentation effectively. Extensive experiments over multiple public datasets show that TPS is simpler to implement, much more stable to train, and achieves superior video segmentation accuracy as compared with the state-of-the-art.

Main Results

SYNTHIA-Seq => Cityscapes-Seq

Methods	road	side.	buil.	pole	light	sign	vege.	sky	per.	rider	car	mIoU
Source	56.3	26.6	75.6	25.5	5.7	15.6	71.0	58.5	41.7	17.1	27.9	38.3
DA-VSN	89.4	31.0	77.4	26.1	9.1	20.4	75.4	74.6	42.9	16.1	82.4	49.5
PixMatch	90.2	49.9	75.1	23.1	17.4	34.2	67.1	49.9	55.8	14.0	84.3	51.0
TPS	91.2	53.7	74.9	24.6	17.9	39.3	68.1	59.7	57.2	20.3	84.5	53.8

VIPER => Cityscapes-Seq

Methods	road	side.	buil.	fence	light	sign	vege.	terr.	sky	per.	car	truck	bus	motor	bike	mIoU
Source	56.7	18.7	78.7	6.0	22.0	15.6	81.6	18.3	80.4	59.9	66.3	4.5	16.8	20.4	10.3	37.1
PixMatch	79.4	26.1	84.6	16.6	28.7	23.0	85.0	30.1	83.7	58.6	75.8	34.2	45.7	16.6	12.4	46.7
DA-VSN	86.8	36.7	83.5	22.9	30.2	27.7	83.6	26.7	80.3	60.0	79.1	20.3	47.2	21.2	11.4	47.8
TPS	82.4	36.9	79.5	9.0	26.3	29.4	78.5	28.2	81.8	61.2	80.2	39.8	40.3	28.5	31.7	48.9

Note: PixMatch is reproduced with replacing the image segmentation backbone to a video segmentaion one.

Installation

create conda environment

conda create -n TPS python=3.6
conda activate TPS
conda install -c menpo opencv
pip install torch==1.2.0 torchvision==0.4.0

clone the ADVENT repo

git clone https://github.com/valeoai/ADVENT
pip install -e ./ADVENT

clone the current repo

git clone https://github.com/xing0047/TPS.git
pip install -r ./TPS/requirements.txt

resample2d dependency:

python ./TPS/tps/utils/resample2d_package/setup.py build
python ./TPS/tps/utils/resample2d_package/setup.py install

Data Preparation

Cityscapes-Seq

TPS/data/Cityscapes/
TPS/data/Cityscapes/leftImg8bit_sequence/
TPS/data/Cityscapes/gtFine/

VIPER

TPS/data/Viper/
TPS/data/Viper/train/img/
TPS/data/Viper/train/cls/

Synthia-Seq

TPS/data/SynthiaSeq/
TPS/data/SynthiaSeq/SEQS-04-DAWN/

Pretrained Models

Download here and put them under pretrained_models.

Optical Flow Estimation

For quick preparation, please download the estimated optical flow of all datasets here.

Synthia-Seq

train
VIPER

train
Cityscapes-Seq

train | val

Train and Test

Train

  cd tps/scripts
  CUDA_VISIBLE_DEVICES=0 python train.py --cfg configs/tps_syn2city.yml
  CUDA_VISIBLE_DEVICES=0 python train.py --cfg configs/tps_viper2city.yml

Test (may in parallel with Train)

  cd tps/scripts
  CUDA_VISIBLE_DEVICES=1 python test.py --cfg configs/tps_syn2city.yml
  CUDA_VISIBLE_DEVICES=1 python test.py --cfg configs/tps_viper2city.yml

Acknowledgement

This codebase is heavily borrowed from DA-VSN.

Contact

If you have any questions, feel free to contact: xing0047@e.ntu.edu.sg or dayan.guan@outlook.com.

Dayan-Guan / TPS