MSANet: Multi-Scale Attention Network for Vertical Seed Distribution in Soybean Breeding Fields

Dataset

Please fill this form to get download link for the datasets and pre-trained weights.

Train

Train the MSANet with our datasets:

1. Download the data.zip following the above Dataset section and unzip the data/ to the root of this repo.
2. Check the options.py.
3. Train the MSANet by running.

python train.py --project_name your_project_name_here --data-root ./data/soypod-200-txt --output_dir ./runs --num_workers 0

Train the MSANet with your own dataset:

First please put you data under "data/" for convenience. (of course you can put your data anywhere else but just remember to modify of provide the '--data_root' item.)

We originally support dataset annotated by V7 (https://www.v7labs.com/). For other formats of the annotation, please modify the get_points function in the utils.py which can read your json format annotation and return the list of [y, x] coordination of point annotations.

Then, train the MSANet by running:

python train.py --project_name your_project_name_here --data-root /path/to/your/dataset --output_dir ./runs --num_workers 0

Evaluation

Statistical Results

We provide evaluation metric for both counting and localization tasks to evaluate our MSANet statistically.

• For couting tasks, we evaluated R2, MAE and RMSE.
• For localization tasks, we evaluated the MED, which is defined in the paper, Precision, Recall and F1 score.

Dataset	R2	MAE	RMSE	MED	Precision	Recall	F1 score
2021 Dataset	0.94	9.20	13.16	7.52	0.87	0.85	0.86
2021 Enlarged Dataset	0.86	13.69	18.32	8.08	0.81	0.87	0.84
2022 Dataset	0.82	13.66	17.26	4.99	0.91	0.85	0.88

We also provide the comparision results with P2PNet-Soy (Jiangsan et al., 2023), please check the jupyter notebooks located under /evaluations/.

Inference

To inference the model, here's an introduction here on Colab.