TasselNetV3

by Hao Lu

Installation

The code has been tested on Python 3.7.4 and PyTorch 1.2.0. Please follow the official instructions to configure your environment. See other required packages in requirements.txt.

Prepare Your Data

Maize Tassels Counting

• Download the Maize Tassels Counting (MTC) dataset from: BaiduYun (1.64 GB) (code: m8rj) or Google Drive (1.8 GB)
• Unzip the dataset and move it into the ./data folder, the path structure should look like this:

$./data/maize_tassels_counting_dataset
├──── trainval
│    ├──── images
│    └──── labels
├──── test
│    ├──── images
│    └──── labels
├──── train.txt
├──── test.txt

Training

Run the following command to train TasselNetv2+ on the MTC dataset:

python --cfg config/mtc-tasselnetv2plus.yaml

• Setting VAL.evaluate_only=False and VAL.visualization=False
• Use CUDA_VISIBLE_DEVICES trick if you have multiple GPUs

Tips and Tricks

If you find some useful tricks and tips, please share it here.

• (Hao Lu) Do not fix bn when training with pretrained models (batch_size=16 tested)
• (Hao Lu) Scale the ground truth by x10 for density-map-based methods when L2 Loss is used (reduction='mean')

Inference

Once the training is finished, run the same command above with VAL.evaluate_only=True for inference.

• Setting VAL.visualization=True to output visualizations. Visualizations are saved in the path ./results/<dataset>/<exp>/<epoch>.

Benchmark Results

Plant Counting

Maize Tassels Counting

Method	Venue, Year	Pretrained	#Param.	MAE	MSE	rMAE	R2	Model
CSRNet	CVPR 2018	VGG16	16.3M	9.43	14.43	100.65	0.7573	One Drive (116MB)
TasselNetv2	PLME 2019	No	525K	5.42	9.21	31.94	0.8923	Baidu Yun (2MB) (code: hrhi)
TasselNetv2+	TBD	No	262K	5.41	9.31	37.65	0.8937	Baidu Yun (2MB) (code: hbnx)
BCNet-BN	TCSVT 2019	VGG16	14.8M	5.11	9.58	27.84	0.8749	Baidu Yun (105MB) (code: mnys)

Maize Tassels Counting (UAV)

Method	Venue, Year	Resolution	Pretrained	#Param.	MAE	MSE	rMAE	R2	Model
TasselNetv2+	TBD	1/8	No	262K	27.08	38.38	14.61	0.8958	-
TasselNetv2+	TBD	1/4	No	262K	16.43	25.79	9.67	0.9515	Baidu Yun (2MB) (code: 68dn)
CSRNet	CVPR 2018	1/4	VGG16	16.3M	14.38	20.52	9.56	0.9704	One Drive (116MB)
BCNet-BN	TCSVT 2019	1/4	VGG16	14.8M	14.37	21.37	8.75	0.9659	Baidu Yun (105MB) (code: t81t)