refac and add more document

app.py

@@ -71,8 +71,7 @@ def query_image(
     pca_features_rem = pca.transform(features[pca_features_fg])
 
     # Min Max Normalization
-    for i in range(3):
-        pca_features_rem[:, i] = (pca_features_rem[:, i] - pca_features_rem[:, i].min()) / (pca_features_rem[:, i].max() - pca_features_rem[:, i].min())
+    pca_features_rem = sklearn.preprocessing.minmax_scale(pca_features_rem)
 
     pca_features_rgb = np.zeros((4 * patch_h * patch_w, 3))
     pca_features_rgb[pca_features_bg] = 0
@@ -82,7 +81,20 @@ def query_image(
     return [pca_features_rgb[i] for i in range(4)]
 
 description = """
-DINOV2 PCA
+DINOV2 PCA demo for <a href="https://arxiv.org/abs/2304.07193">DINOv2: Learning Robust Visual Features without Supervision(Figure 1)</a>
+
+How to Use:
+
+1. Enter 4 images that have clean background and similar object.
+2. Edit threshold and checkbox to split background/foreground.
+
+Method:
+1. Compute the features of patches from 4 images. We can get a feature that have (4 * patch_w * patch_h, feature_dim) shape.
+2. PCA the feature with 3 dims. After PCA, Min-Max normalization is performed.
+3. Use first component to split foreground and background. (threshold and checkbox)
+4. All the feature of patches included in the background are set to 0.=
+5. PCA is performed based on the remaining features. Afer PCA, Min-Max normalization is performed.
+6. Visualize
 """
 demo = gr.Interface(
     query_image,