evaluation: resize segmentation input with OpenCV to reduce difference with Android results

evaluation/python/generate_masks.py

@@ -2,7 +2,7 @@
 
 import numpy as np
 import tensorflow as tf
-from PIL import Image, ImageOps
+import cv2
 from pathlib import Path
 
 SEG_MODEL_FILE_PATH = "../../app/build/downloads/fairscan-segmentation-model.tflite"
@@ -12,15 +12,21 @@ INPUT_WIDTH = 256
 INPUT_HEIGHT = 256
 
 def get_resized_image(image_path):
-    img = Image.open(image_path).convert("RGB")
-    img = ImageOps.exif_transpose(img)
-    img = img.convert("RGB").resize((INPUT_WIDTH, INPUT_HEIGHT), Image.BILINEAR)
+    img = cv2.imread(str(image_path), cv2.IMREAD_COLOR)
+    if img is None:
+        raise ValueError(f"Failed to load image: {image_path}")
+    img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
+    img = cv2.resize(
+        img,
+        (INPUT_WIDTH, INPUT_HEIGHT),
+        interpolation=cv2.INTER_LINEAR
+    )
     return img
 
-def preprocess_image(img: Image.Image) -> np.ndarray:
-    img_np = np.asarray(img).astype(np.float32)
-    img_np = (img_np - 127.5) / 127.5  # Normalize to [-1, 1]
-    return np.expand_dims(img_np, axis=0)
+def preprocess_image(img):
+    img = img.astype(np.float32)
+    img = img / 127.5 - 1.0
+    return img[np.newaxis, ...]
 
 def postprocess_output(output: np.ndarray) -> np.ndarray:
     output = np.squeeze(output).astype(np.float32)  # Shape: (256, 256)
@@ -45,7 +51,7 @@ mask_input_dir = DATASET_DIR / "masks"
 for image_path in sorted(img_input_dir.glob("*.jpg")):
     print(f"Generating mask for {image_path}")
     img = get_resized_image(image_path)
-    img = ImageOps.exif_transpose(img)
     mask = get_segmentation_mask(img)
     mask_path = mask_input_dir / (image_path.stem + ".png")
-    Image.fromarray((mask * 255).astype(np.uint8)).save(mask_path)
+    mask_uint8 = (mask * 255.0).round().astype(np.uint8)
+    cv2.imwrite(str(mask_path), mask_uint8)

evaluation/python/requirements.txt

@@ -1,3 +1,3 @@
 tensorflow
 numpy
-Pillow
+opencv-python