类别无关植物计数(CACOP)项目进展

---

数据集收集

• 一千张植物图片已收集标注完毕，并整理完数据集类别表；

• 生成了数据集相关的注释和划分文件、密度图，项目数据集结构如下：

  our_data
  ├── Train_Val_Test.json
  ├── annotations.json
  ├── gt_density_maps
  └── images   
  

• 数据集已上传至谷歌网盘

---

模型复现

已成功复现5个模型，模型复现结果如下：

• BMNet+

  train：MAE: 8.63, rMAE: 0.58, RMSE: 12.36, rRMSE: 0.95, R2: 0.59, FLOPS: 83.8G
  val：  MAE: 7.51, rMAE: 0.43, RMSE: 11.00, rRMSE: 0.65, R2: 0.57, FLOPS: 78.7G
  test： MAE: 6.47, rMAE: 0.50, RMSE: 12.00, rRMSE: 0.71, R2: 0.87, FLOPS: 68.4G
  

• CACViT

  [23:14:23.329991]Averaged stats:
  [23:14:23.330031]MAE:5.25,RMSE:7.37,rMAE:0.42,rRMSE:0.44,R2:0.95
  [23:14:23.330043]
  FLOPs Summary:   失败的总结:
  [23:14:23.330057]Total FLOPs for all samples: 67.9290 TFLOPs[23:14:23.330057]所有样品的总FLOPs: 67.9290 TFLOPs
  [23:14:23.330069]Average FLOPs per sample:449.8607 GFLOPs
  [23:14:23.330089]Min sample FLOPs:177.5921 GFLOPs
  [23:14:23.330102]MaX Sample FLOPs:1509.5326 GFLOPs
  [23:14:23.330367]Median sample FLOPs:443.9802 GFLOPs
  [23:14:23.330393]Testing time 0:00:57
  

• FamNet

  lr=1e-5：
  train：MAE: 9.07 , rMAE: 0.4213, RMSE: 13.21, rRMSE: 0.6139, R²: 0.5318, FLOPS: 133.4948G
  val：  MAE: 10.02, rMAE: 0.4845, RMSE: 15.03, rRMSE: 0.7269, R²: 0.1955, FLOPS: 133.4948G
  test： MAE: 13.41, rMAE: 0.8094, RMSE: 33.31, rRMSE: 2.0101, R²: 0.0251, FLOPS: 133.4948G
  

• SPDCN

  [2025-06-14 15:00:12 SPDCN](val.py 176): INFO  * MAE 10.505 RMSE 15.547
  [2025-06-14 15:00:12 SPDCN](val.py 177): INFO  * rMAE 0.643 rRMSE 0.754
  [2025-06-14 15:00:12 SPDCN](val.py 178): INFO  * R2 0.759
  [2025-06-14 15:00:12 SPDCN](val.py 180): INFO  * Average FLOPs per sample: 70.45G
  [2025-06-14 15:00:12 SPDCN](val.py 181): INFO  * Median sample FLOPs: 563.61G
  

• GeCo

  Train：MAE: 10.53, rMAE: 0.4893, RMSE: 16.42, rRMSE: 0.7630, R2: 0.2768, FLOPS: 5712.09G
  Val：  MAE: 6.77 , rMAE: 0.3275, RMSE: 9.10 , rRMSE: 0.4403, R2: 0.7049, FLOPS: 5712.09G
  Test： MAE: 6.74 , rMAE: 0.4069, RMSE: 10.56, rRMSE: 0.6375, R2: 0.9019, FLOPS: 5712.09G
  

Demo

• 代码见\Demo，基于flask实现，运行后在浏览器打开生成的网址
• 基于可视化效果最优的GeCo实现
• 由于GeCo显存占用大，因此还开发了基于CACViT的轻便版本（demo_CACVIT.py、predictor_CACVIT.py、\templates\demo_CACVIT.html）
• 示例./demo.mp4

demo.mp4