PointFix: Learning to Fix Domain Bias for Robust Online Stereo Adaptation(PointFix) ECCV 2022

Kwonyoung Kim¹ Jungin Park¹ Jiyoung Lee² Dongbo Min³ Kwanghoon Sohn¹

¹_{Yonsei University}
²_{NAVER AI LAB}
³_{Ewha University}

Required packages can be found in the packagelist.txt.

We start PointFix training from the pretrained initial weights and they can be found in the Real-Time Self-Adaptive Deep Stereo page.

MADNet parameters pretrained by PointFix are available here.

Example of PointFix training for MADNet:

OUTPUT="path/to/output/folder"
DATASET="./example_dataset.csv"
BATCH_SIZE="4"
ITERATIONS=30000
PRETRAINED_WEIGHTS="./pretrained_nets/MADNet/synthetic/weights.ckpt"
MODEL_NAME="MADNet"
LR="0.0001"
ALPHA="0.00001"
LOSS="mean_l1"
ADAPTATION_LOSS="mean_SSIM_l1"
META_ALGORITHM="PointFix"
RESIZE_SHAPE="380 640"
DATASET_PARAM="Synthia"

python train_pointfix.py --dataset $DATASET -o $OUT_FOLDER -b $BATCH_SIZE -n $ITERATIONS --adaptationSteps $ADAPTATION_ITERATION \
--weights $PRETRAINED_WEIGHTS --lr $LR --alpha $ALPHA --loss $LOSS --adaptationLoss $ADAPTATION_LOSS --unSupervisedMeta \
--metaAlgorithm $META_ALGORITHM --resizeShape $RESIZE_SHAPE --dataset_param $DATASET_PARAM \
--modelName $MODEL_NAME

Example of online adaptation test using MADNet and MAD:

LIST="./example_sequence.csv" 
OUTPUT="path/to/output/folder"
WEIGHTS="path/to/pretrained/network"
MODELNAME="MADNet"
BLOCKCONFIG="block_config/MadNet_full.json"
MODE="MAD"
LR=0.00001

python3 test_online_adaptation.py \
    -l ${LIST} \
    -o ${OUTPUT} \
    --weights ${WEIGHTS} \
    --modelName MADNet \
    --blockConfig ${BLOCKCONFIG} \
    --mode ${MODE} \
    --sampleMode PROBABILITY \
    --logDispStep 1 \
    --lr ${LR}