[experiment, SDK] Run inference on selected ROI

gen2-inference-on-roi/main.py

@@ -0,0 +1,84 @@
+import cv2
+from depthai_sdk import OakCamera
+from depthai_sdk.classes.packets import DetectionPacket
+import numpy as np
+import depthai as dai
+
+roi = None
+new_roi = False
+selecting_roi = False
+def roi_selection(event, x, y, flags, param):
+    global start_pts, selecting_roi, roi, new_roi
+
+    if event == cv2.EVENT_LBUTTONDOWN:
+        selecting_roi = True
+        start_pts = [x, y]
+
+    elif event == cv2.EVENT_MOUSEMOVE:
+        if selecting_roi:
+            start_pts[2:] = [x, y]
+            roi = start_pts
+
+    elif event == cv2.EVENT_LBUTTONUP:
+        selecting_roi = False
+        new_roi = True
+        if abs(start_pts[0] - x) < 100 or abs(start_pts[1] - y) < 100:
+            roi = None
+            return
+        start_pts[2:] = [x, y]
+        roi = start_pts
+
+
+shape = None
+def cb(packet: DetectionPacket):
+    global roi, shape
+    frame = packet.visualizer.draw(packet.frame)
+    if roi is not None:
+        x_start, y_start, x_end, y_end = roi
+        # frame_roi = frame[y_start:y_end, x_start:x_end]
+
+        # Create a mask with the same size as the frame
+        mask = np.zeros_like(frame)
+        mask[y_start:y_end, x_start:x_end] = frame[y_start:y_end, x_start:x_end]
+
+        # Darken the unselected region
+        frame_darkened = frame * 0.3
+        frame_darkened = frame_darkened.astype(np.uint8)
+
+        # Blend the mask with the darkened frame
+        frame = cv2.addWeighted(frame_darkened, 1, mask, 1, 0)
+    shape = frame.shape
+    cv2.imshow('color', frame)
+
+with OakCamera() as oak:
+    cv2.namedWindow("color")
+    cv2.setMouseCallback("color", roi_selection)
+
+    color = oak.create_camera('color')
+    nn = oak.create_nn('mobilenet-ssd', color)
+    oak.visualize(color, fps=True, callback=cb)
+    oak.visualize([nn.out.passthrough, nn], fps=True, scale=2/3)
+
+    manipIn = oak.pipeline.create(dai.node.XLinkIn)
+    manipIn.setStreamName('manip_in')
+    manipIn.out.link(nn.image_manip.inputConfig)
+
+    oak.start() # Initialize the device and start the pipeline
+
+    raw_cfg = nn.image_manip.initialConfig.get() # To keep existing settings
+
+    manipQ = oak.device.getInputQueue('manip_in')
+
+    while oak.running():
+        oak.poll()
+        if new_roi:
+            new_roi = False
+            conf = dai.ImageManipConfig()
+            conf.set(raw_cfg)
+            if roi is not None:
+                normalized_roi = (roi[0]/shape[1], roi[1]/shape[0], roi[2]/shape[1], roi[3]/shape[0])
+                conf.setCropRect(normalized_roi)
+            manipQ.send(conf)
+            print(roi)
+
+cv2.destroyAllWindows()