[WIP]: Point detection model solves #809

deepforest/convert centroids.py

@@ -0,0 +1,11 @@
+import pandas as pd
+
+df = pd.read_csv('deepforest/data/OSBS_029.csv')
+
+# Calculate centroids
+df['x_center'] = round((df['xmin'] + df['xmax']) / 2)
+df['y_center'] = round((df['ymin'] + df['ymax']) / 2)
+
+# Drop original bounding box columns
+df_centroids = df.drop(columns=["xmin", "ymin", "xmax", "ymax"])
+df_centroids.to_csv("deepforest/data/OSBS_029_centroids.csv")

deepforest/data/OSBS_029_centroids.csv

@@ -0,0 +1,62 @@
+,image_path,label,x_center,y_center
+0,OSBS_029.tif,Tree,215.0,78.0
+1,OSBS_029.tif,Tree,272.0,120.0
+2,OSBS_029.tif,Tree,196.0,278.0
+3,OSBS_029.tif,Tree,382.0,14.0
+4,OSBS_029.tif,Tree,330.0,30.0
+5,OSBS_029.tif,Tree,382.0,46.0
+6,OSBS_029.tif,Tree,295.0,19.0
+7,OSBS_029.tif,Tree,382.0,225.0
+8,OSBS_029.tif,Tree,106.0,131.0
+9,OSBS_029.tif,Tree,132.0,130.0
+10,OSBS_029.tif,Tree,180.0,173.0
+11,OSBS_029.tif,Tree,140.0,172.0
+12,OSBS_029.tif,Tree,362.0,305.0
+13,OSBS_029.tif,Tree,27.0,185.0
+14,OSBS_029.tif,Tree,21.0,236.0
+15,OSBS_029.tif,Tree,91.0,166.0
+16,OSBS_029.tif,Tree,384.0,94.0
+17,OSBS_029.tif,Tree,176.0,126.0
+18,OSBS_029.tif,Tree,172.0,212.0
+19,OSBS_029.tif,Tree,128.0,220.0
+20,OSBS_029.tif,Tree,62.0,388.0
+21,OSBS_029.tif,Tree,134.0,384.0
+22,OSBS_029.tif,Tree,244.0,255.0
+23,OSBS_029.tif,Tree,314.0,384.0
+24,OSBS_029.tif,Tree,354.0,360.0
+25,OSBS_029.tif,Tree,20.0,38.0
+26,OSBS_029.tif,Tree,20.0,86.0
+27,OSBS_029.tif,Tree,76.0,30.0
+28,OSBS_029.tif,Tree,116.0,52.0
+29,OSBS_029.tif,Tree,168.0,22.0
+30,OSBS_029.tif,Tree,208.0,20.0
+31,OSBS_029.tif,Tree,391.0,282.0
+32,OSBS_029.tif,Tree,347.0,144.0
+33,OSBS_029.tif,Tree,346.0,101.0
+34,OSBS_029.tif,Tree,378.0,130.0
+35,OSBS_029.tif,Tree,133.0,285.0
+36,OSBS_029.tif,Tree,128.0,340.0
+37,OSBS_029.tif,Tree,196.0,378.0
+38,OSBS_029.tif,Tree,215.0,357.0
+39,OSBS_029.tif,Tree,196.0,326.0
+40,OSBS_029.tif,Tree,259.0,324.0
+41,OSBS_029.tif,Tree,292.0,353.0
+42,OSBS_029.tif,Tree,72.0,215.0
+43,OSBS_029.tif,Tree,133.0,84.0
+44,OSBS_029.tif,Tree,74.0,93.0
+45,OSBS_029.tif,Tree,276.0,262.0
+46,OSBS_029.tif,Tree,350.0,64.0
+47,OSBS_029.tif,Tree,268.0,67.0
+48,OSBS_029.tif,Tree,312.0,114.0
+49,OSBS_029.tif,Tree,302.0,152.0
+50,OSBS_029.tif,Tree,225.0,114.0
+51,OSBS_029.tif,Tree,273.0,215.0
+52,OSBS_029.tif,Tree,44.0,356.0
+53,OSBS_029.tif,Tree,36.0,384.0
+54,OSBS_029.tif,Tree,16.0,278.0
+55,OSBS_029.tif,Tree,93.0,264.0
+56,OSBS_029.tif,Tree,78.0,312.0
+57,OSBS_029.tif,Tree,102.0,376.0
+58,OSBS_029.tif,Tree,244.0,142.0
+59,OSBS_029.tif,Tree,331.0,194.0
+60,OSBS_029.tif,Tree,236.0,226.0

deepforest/dataset.py

@@ -248,3 +248,49 @@ def __getitem__(self, idx):
             image = box
 
         return image
+
+
+class PointDataset(Dataset):
+    """An in-memory dataset for point predictions (centroids).
+
+    Args:
+        df: a pandas dataframe with image_path, x_center, y_center, and label columns.
+        transform: a function to apply to the image.
+        root_dir: the directory where the image is stored.
+
+    Returns:
+        rgb: a tensor of shape (3, height, width)
+    """
+
+    def __init__(self, df, root_dir, transform=None, augment=False):
+        self.df = df
+        self.root_dir = root_dir
+        self.transform = transform
+        self.augment = augment
+
+        unique_image = self.df['image_path'].unique()
+        assert len(unique_image
+                  ) == 1, "There should be only one unique image for this class object"
+
+        # Open the image using rasterio
+        self.src = rio.open(os.path.join(root_dir, unique_image[0]))
+
+    def __len__(self):
+        return len(self.df)
+
+    def __getitem__(self, idx):
+        row = self.df.iloc[idx]
+        x_center = row['x_center']
+        y_center = row['y_center']
+
+        # Read a small window around the centroid to create a context image for the point (optional)
+        box_size = 20  # e.g., small crop around the centroid
+        window = Window(x_center - box_size // 2, y_center - box_size // 2, box_size,
+                        box_size)
+        image = self.src.read(window=window)
+        image = np.rollaxis(image, 0, 3)
+
+        if self.transform:
+            image = self.transform(image)
+
+        return image, (x_center, y_center)