TasselNetV4: A vision foundation model for cross-scene, cross-scale, and cross-species plant counting

Official implementation of TasselNetV4

ISPRS Journal of Photogrammetry and Remote Sensing (IF=12.2).

Xiaonan Hu, Xuebing Li, Jinyu Xu, Abdulkadir Duran Adan, Xuhui Zhu, Yanan Li, Wei Guo, Shouyang Liu, Wenzhong Liu, Hao Lu*

Highlights

• Plant agnostic counting: a new plant-orientated task customizing Class-agnostic counting into the plant domain and highlighting zero-shot generalization across taxomomic plant species;
• PAC-105 and PAC-Somalia: two challenging PAC datasets for training and evaluating daily and out-of-distribution plant species;
• TasselNetV4: an extended version of the TasselNet plant counting models.

Comparison with the state-of-the-art CAC approaches on the PAC-105 dataset. Best performance is in boldface.

Method	Venue & Year	Shot	MAE↓	RMSE↓	WCA↑	$R^2$ ↑	checkpoints
FamNet	CVPR'21	3	31.70	62.58	0.49	0.56	Baiduyun
BMNet+	CVPR'22	3	27.03	60.18	0.56	0.61	Baiduyun
SPDCNet	BMVC'22	3	25.21	49.98	0.58	0.92	Baiduyun
SAFECount	WACV'23	3	25.59	52.09	0.58	0.91	Baiduyun
CountTR	BMVC'22	3	25.25	49.31	0.63	0.92	Baiduyun
T-Rex2	ECCV'24	3	26.04	49.31	0.58	0.92
CACViT	AAAI'24	3	19.51	29.59	0.68	0.89	Baiduyun
TasselNetV4 (Ours)	ISPRS'25	3	16.04	28.03	0.74	0.92
FamNet	CVPR'21	1	35.91±0.966	71.78±1.188	0.42±0.014	0.45±0.024
BMNet	CVPR'22	1	28.78±0.324	62.12±0.437	0.15±0.001	0.59±0.008
CountTR	BMVC'22	1	28.46±0.226	49.84±0.646	0.70±0.037	0.73±0.006
CACViT	AAAI'24	1	21.80±0.429	38.40±1.526	0.64±0.005	0.84±0.013
TasselNetV4 (Ours)	ISPRS'25	1	18.04±0.339	32.04±1.213	0.71±0.005	0.90±0.009

Comparison with the state-of-the-art CAC approaches on the PAC-Somalia dataset. Best performance is in boldface.

Method	Venue & Year	Shot	MAE↓	RMSE↓	WCA↑	$R^2$ ↑
CountTR	BMVC'22	3	12.71	23.87	0.38	0.57
CACViT	AAAI'24	3	14.00	17.00	0.55	0.78
TasselNetV4 (Ours)	This Paper	3	8.88	13.11	0.72	0.87
CountTR	BMVC'22	1	12.79±0.076	24.20±0.292	0.37±0.005	0.55±0.012
CACViT	AAAI'24	1	14.74±0.138	18.23±0.446	0.53±0.005	0.75±0.012
TasselNetV4 (Ours)	ISPRS'25	1	10.98±0.065	16.73±0.150	0.65±0.000	0.80±0.005

Visualization

Installation

To setup all the required dependencies for training and evaluation, please follow the instructions below:

conda env create -f environment.yaml
conda activate TN4

Prepare Dataset

PAC-105&PAC-Somalia

Download training dataset PAC-105 from Baiduyun (2.8G) | Google Drive (2.8G) and test dataset PAC-Somalia from: Baiduyun (208M) | Google Drive (208M). The dataset structure should look like this:

/dataset
├──── aska
    ├──── aska_fruit
        ├──── images
            ├──── image1.png
            └──── ...
        └──── labels
            └──── aska.csv
├──── boocbooc
├──── ......
├──── yicib
└──── dataset.csv

Inference

First modify Dataset.data_path and Resume.resume_path in cfg_test/local_32_64_128_loose.yml;

Download our model from Baiduyun (1.05G) | Google Drive (1.05G);

Run the following command to reproduce our results of TasselNetV4 on the PAC-105 / PAC-Somalia:

python test_checkpoint.py

• Results will be saved in the path ./visual.

Training

First modify path in cfg_train/local_32_64_128_loose.yml;

Run the following command to train your model

python train_val.py

License

This repository is released under the Apache 2.0 license as found in the LICENSE.