Mediapipe `max_num_hands` not working properly

[Duy Khanh]

I am trying to extract hand pose using Mediapipe. Even though `max_num_hands` is set to be 2, sometimes `results.multi_hand_landmarks` still returns an array of length 3, which indicates 3 hands. I checked this by `print(len(results.multi_hand_landmarks))`.

The function I use to extract pose (the commented code is for displaying):

import cv2 

import mediapipe as mp 

mp_hands = mp.solutions.hands 

hands = mp_hands.Hands(

        static_image_mode=False, 

        max_num_hands=2, 

        min_detection_confidence=0.5, 

        min_tracking_confidence=0.5, ) 

def extract_pose(frame, hands): 

        frame = cv2.flip(frame, 1) 

        frame_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB) 

        results = hands.process(frame_rgb) 

        if results.multi_hand_landmarks: 

                print(len(results.multi_hand_landmarks)) 

                # for hand_landmarks in results.multi_hand_landmarks: 

                # for landmark in hand_landmarks.landmark: 

                # x = int(landmark.x * frame.shape[1]) 

                # y = int(landmark.y * frame.shape[0]) 

                # 

                # cv2.circle(frame, (x, y), 5, (0, 255, 0), -1) 

                # cv2.imshow("Hand Pose", frame)

And the code to read images:
         for idx in range(960, 964): 

                path = f'data/frame/s1_Jeremy/s1_Jeremy_2/s1_Jeremy_2_{idx:04d}.jpg' 

                print(path) img = cv2.imread(path, cv2.IMREAD_COLOR) 

                extract_pose(img, hands) 

                cv2.waitKey(0) 

                cv2.destroyAllWindows()

For some reason, this only happens if I read frames continuously (as in the code to read images above). The extra pose detected is very close to the existing hand pose.

How do I force mediapipe to extract only 2 poses at maximum, and of different hands, so that it does not extract pose from the same hand twice?