Add the quantitative comparisons of BiRefNet_dynamics and previous BiRefNet models for genenral use.

README.md

@@ -19,6 +19,19 @@ license: mit
 > An arbitrary shape adaptable BiRefNet for general segmentation.  
 > This model was trained on arbitrary shapes (256x256 ~ 2304x2304) and shows great robustness on inputs with any shape.
 
+### Performance
+> How it looks when compared with BiRefNet-general (fixed 1024x1024 resolution) -- greater than BiRefNet-general and BiRefNet_HR-general on the reserved validation sets (DIS-VD and TE-P3M-500-NP).  
+> The `dynamic_XXxXX` means this BiRefNet_dynamic model was being tested in various input resolutions for the evaluation.
+
+
+![dynamic_shapes-comparison](https://cdn-uploads.huggingface.co/production/uploads/63e4977454f51ea342d54814/TrV23TQIutQJCor0T9mYY.png)
+
+![comparison_DIS-VD](https://cdn-uploads.huggingface.co/production/uploads/63e4977454f51ea342d54814/swpfpyeFVRf3aUUj7UjOD.png)
+
+![comparison_TE-P3M-500-NP](https://cdn-uploads.huggingface.co/production/uploads/63e4977454f51ea342d54814/k9pypZIvPvM1UC0tsMals.png)
+
+For performance of different epochs, check the [eval_results-xxx folder for it](https://drive.google.com/drive/u/0/folders/1J79uL4xBaT3uct-tYtWZHKS2SoVE2cqu) on my google drive.
+
 <div align='center'>
     <a href='https://scholar.google.com/citations?user=TZRzWOsAAAAJ' target='_blank'><strong>Peng Zheng</strong></a><sup> 1,4,5,6</sup>,&thinsp;
     <a href='https://scholar.google.com/citations?user=0uPb8MMAAAAJ' target='_blank'><strong>Dehong Gao</strong></a><sup> 2</sup>,&thinsp;