@ARTICLE {ZoomNeXt,
title = {ZoomNeXt: A Unified Collaborative Pyramid Network for Camouflaged Object Detection},
author ={Youwei Pang and Xiaoqi Zhao and Tian-Zhu Xiang and Lihe Zhang and Huchuan Lu},
journal = {IEEE Transactions on Pattern Analysis and Machine Intelligence},
year = {2024},
doi = {10.1109/TPAMI.2024.3417329},
}
Datasets | Links |
---|---|
CAMO-TE, CHAMELEON, COD10K-TE, NC4K | ResNet-50, EfficientNet-B4, PVTv2-B2, PVTv2-B3, PVTv2-B4, PVTv2-B5 |
CAD, MoCA-Mask-TE | PVTv2-B5 |
Backbone | CAMO-TE | CHAMELEON | COD10K-TE | NC4K | Links | ||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|
$S_m$ | $F^{\omega}_{\beta}$ | MAE | $S_m$ | $F^{\omega}_{\beta}$ | MAE | $S_m$ | $F^{\omega}_{\beta}$ | MAE | $S_m$ | $F^{\omega}_{\beta}$ | MAE | ||
ResNet-50 | 0.833 | 0.774 | 0.065 | 0.908 | 0.858 | 0.021 | 0.861 | 0.768 | 0.026 | 0.874 | 0.816 | 0.037 | Weight |
EfficientNet-B1 | 0.848 | 0.803 | 0.056 | 0.916 | 0.870 | 0.020 | 0.863 | 0.773 | 0.024 | 0.876 | 0.823 | 0.036 | Weight |
EfficientNet-B4 | 0.867 | 0.824 | 0.046 | 0.911 | 0.865 | 0.020 | 0.875 | 0.797 | 0.021 | 0.884 | 0.837 | 0.032 | Weight |
PVTv2-B2 | 0.874 | 0.839 | 0.047 | 0.922 | 0.884 | 0.017 | 0.887 | 0.818 | 0.019 | 0.892 | 0.852 | 0.030 | Weight |
PVTv2-B3 | 0.885 | 0.854 | 0.042 | 0.927 | 0.898 | 0.017 | 0.895 | 0.829 | 0.018 | 0.900 | 0.861 | 0.028 | Weight |
PVTv2-B4 | 0.888 | 0.859 | 0.040 | 0.925 | 0.897 | 0.016 | 0.898 | 0.838 | 0.017 | 0.900 | 0.865 | 0.028 | Weight |
PVTv2-B5 | 0.889 | 0.857 | 0.041 | 0.924 | 0.885 | 0.018 | 0.898 | 0.827 | 0.018 | 0.903 | 0.863 | 0.028 | Weight |
Backbone | CAD | MoCA-Mask-TE | Links | ||||||||
---|---|---|---|---|---|---|---|---|---|---|---|
$S_m$ | $F^{\omega}_{\beta}$ | MAE | mDice | mIoU | $S_m$ | $F^{\omega}_{\beta}$ | MAE | mDice | mIoU | ||
PVTv2-B5 (T=5) | 0.757 | 0.593 | 0.020 | 0.599 | 0.510 | 0.734 | 0.476 | 0.010 | 0.497 | 0.422 | Weight |
Set all dataset information to the dataset.yaml
as follows.
pip install -r requirements.txt
# ICOD
python main_for_image.py --config configs/icod_train.py --model-name <MODEL_NAME> --evaluate --load-from <TRAINED_WEIGHT>
# VCOD
python main_for_video.py --config configs/vcod_finetune.py --model-name <MODEL_NAME> --evaluate --load-from <TRAINED_WEIGHT>
[!note]
Evaluating performance on the VCOD dataset directly using training scripts is not consistent with the paper. This is because the evaluation approach in the paper continues the strategy of previous work SLT-Net, which adjusts the range of valid frames in the sequence.
To get the results in our paper, you can use PySODEvalToolkit and use the similar command as:
python ./eval.py `
--dataset-json vcod-datasets.json `
--method-json vcod-methods.json `
--include-datasets CAD `
--include-methods videoPvtV2B5_ZoomNeXt `
--data-type video `
--valid-frame-start "0" `
--valid-frame-end "0" `
--metric-names "sm" "wfm" "mae" "fmeasure" "em" "dice" "iou"
python ./eval.py `
--dataset-json vcod-datasets.json `
--method-json vcod-methods.json `
--include-datasets MOCA-MASK-TE `
--include-methods videoPvtV2B5_ZoomNeXt `
--data-type video `
--valid-frame-start "0" `
--valid-frame-end "-2" `
--metric-names "sm" "wfm" "mae" "fmeasure" "em" "dice" "iou"
python main_for_image.py --config configs/icod_train.py --pretrained --model-name EffB1_ZoomNeXt
python main_for_image.py --config configs/icod_train.py --pretrained --model-name EffB4_ZoomNeXt
python main_for_image.py --config configs/icod_train.py --pretrained --model-name PvtV2B2_ZoomNeXt
python main_for_image.py --config configs/icod_train.py --pretrained --model-name PvtV2B3_ZoomNeXt
python main_for_image.py --config configs/icod_train.py --pretrained --model-name PvtV2B4_ZoomNeXt
python main_for_image.py --config configs/icod_train.py --pretrained --model-name PvtV2B5_ZoomNeXt
python main_for_image.py --config configs/icod_train.py --pretrained --model-name RN50_ZoomNeXt
python main_for_image.py --config configs/icod_pretrain.py --info pretrain --model-name PvtV2B5_ZoomNeXt --pretrained
python main_for_video.py --config configs/vcod_finetune.py --info finetune --model-name videoPvtV2B5_ZoomNeXt --load-from <PRETAINED_WEIGHT>