readme.md

MINet

Folders & Files

• base: Store some basic code for building the network here.
• loss: The code of the loss function.
• models: The code of important modules.
• network: The code of the network.
• utils: Some instrumental code.
  • utils/config.py: Configuration file for model training and testing.
  • utils/imgs/*py: Some files about creating the dataloader.
• main.py: I think you can understand.

My Environment

• Python=3.7
• PyTorch=1.3.1 (I think >=1.1 is OK. You can try it, please create an issue if there is a problem 😄.)
• tqdm
• prefetch-generator
• tensorboardX

Recommended way to install these packages:

# create env
conda create -n pt13 python=3.7
conda activate pt13

# install pytorch
conda install pytorch=1.3.1 torchvision cudatoolkit=9.0 cudnn -c pytorch

# some tools
pip install tqdm
# (optional) https://github.com/Lyken17/pytorch-OpCounter
# pip install thop
pip install prefetch-generator

# install tensorboard
pip install tensorboardX
pip install tensorflow==1.13.1

Train

1. You can customize the value of the arg_config dictionary in the configuration file.
   • The first time you use it, you need to adjust the path of every dataset.
   • And other setting in config.py, like epoch_num, lr and so on...
2. In the folder code, run the command python main.py.
3. Everything is OK. Just wait for the results.
4. The test will be performed automatically when the training is completed.
5. All results will be saved into the folder output, including predictions in folder pre (if you set save_pre to True), .pth files in folder pth and other log files.

If you want to test the trained model again...

Our pre-training parameters can also be used in this way.

1. In the output folder, please ensure that there is a folder corresponding to the model, which contains the pth folder, and the .pth file of the model is located here.
2. Set the value of NET of arg_config to the model you want to test like these lines.
3. Set the value of epoch_num to 0.
4. In the folder code, run python main.py.
5. You can find predictions from the model in the folder pre of the output.

Evaluation

We evaluate results of all models by lartpang/sal_eval_toolbox.

lartpang/sal_eval_toolbox is based on ArcherFMY/sal_eval_toolbox.

But, we add the code about E-measure and weighted F-measure and update the related code in our forked repository lartpang/sal_eval_toolbox. Welcome to use it in your code ⭐!

More

If there are other issues, you can create a new issue.

More Experiments

F3Net is the most recent work on SOD, and the performance is very good. I think its training strategy is of great reference value. Here, I have changed our training method by learning from its code.

To explore the upper limit of the performance of the model, I tried some ways to improve the performance of the model on a NVIDIA GTX 1080Ti (~11G).

• To achieve a larger batch size:
  • we reduce the number of intermediate channels in AIMs;
  • we apply the checkpoint feature of PyTroch;
    • https://blog.csdn.net/one_six_mix/article/details/93937091
    • https://pytorch.org/docs/stable/checkpoint.html#torch.utils.checkpoint.checkpoint
  • we set the batch size to 32.
• Use more effective training strategies:
  • We apply multi-scale training strategy (borrowed from the code of F3Net).
  • Network parameters are trained in groups with different learning rates (borrowed from the code of F3Net).
  • Multiple learning rate decay strategies with/without the warmup technology.

Results:

D			DO			HI			E			PS			SOC			MT	LRDecay	Optimizer	InitLR	Scale	EP
MAXF	MEANF	MAE	MAXF	MEANF	MAE	MAXF	MEANF	MAE	MAXF	MEANF	MAE	MAXF	MEANF	MAE	MAXF	MEANF	MAE
0.853	0.787	0.048	0.794	0.734	0.060	0.922	0.891	0.036	0.931	0.908	0.043	0.856	0.810	0.084	0.377	0.342	0.086	FALSE	Poly	Sgd_trick	0.05	320	40
0.866	0.793	0.043	0.789	0.722	0.059	0.925	0.888	0.034	0.935	0.905	0.041	0.874	0.822	0.070	0.382	0.347	0.110	FALSE	Poly	Sgd_trick	0.001	320	40
0.881	0.822	0.037	0.803	0.746	0.053	0.934	0.904	0.029	0.942	0.919	0.036	0.880	0.837	0.066	0.390	0.356	0.081	FALSE	Poly	Sgd_trick	0.005	320	40
0.878	0.815	0.039	0.803	0.745	0.054	0.934	0.904	0.029	0.944	0.919	0.035	0.878	0.833	0.067	0.385	0.352	0.079	FALSE	Cos_warmup	Sgd_trick	0.005	320	40
0.878	0.815	0.038	0.797	0.741	0.054	0.931	0.901	0.031	0.941	0.917	0.038	0.875	0.831	0.067	0.382	0.355	0.085	FALSE	Cos_warmup	Sgd_trick	0.003	320	40
0.892	0.836	0.036	0.820	0.763	0.053	0.943	0.918	0.026	0.950	0.929	0.034	0.884	0.847	0.064	0.388	0.355	0.087	TRUE	f3_sche	f3_trick	0.05	352	40
0.891	0.834	0.037	0.820	0.762	0.055	0.942	0.915	0.026	0.948	0.928	0.034	0.888	0.844	0.064	0.394	0.359	0.120	TRUE	Cos_warmup	f3_trick	0.05	352	40
0.895	0.840	0.035	0.816	0.762	0.055	0.942	0.915	0.027	0.947	0.927	0.034	0.884	0.843	0.066	0.395	0.359	0.112	TRUE	Cos_w/o_warmup	f3_trick	0.05	352	40
0.893	0.838	0.036	0.814	0.759	0.056	0.943	0.917	0.026	0.949	0.930	0.033	0.886	0.849	0.065	0.395	0.359	0.134	TRUE	Poly	f3_trick	0.05	352	40

• D: DUTS
• DO: DUT-OMRON
• HI: HKU-IS
• E: ECSSD
• PS: PASCAL-S
• MT: Multi-scale Training
• EP: Epoch Number

NOTE: The results here are for reference only. Note that the results here are all tested on the complete test datasets. In fact, some of the results here can be higher if testing in the way of the existing papers. Because the test set in my paper follows the settings of the existing papers, some datasets, such as HKU-IS, are not tested with the complete dataset.

注： 此处结果仅供参考。请注意，这里的结果都是在完整的测试数据集上测试的。事实上，如果按照现有论文的方式进行测试，这里的一些结果可能会更高。由于本文中的测试集遵循现有论文的设置，一些数据集，如HKU-IS，没有使用完整的数据集进行测试。