def __init__()

in data/InterHand2.6M/dataset.py [0:0]


    def __init__(self, transform, mode):
        self.mode = mode # train, test, val
        self.img_path = '../data/InterHand2.6M/images'
        self.annot_path = '../data/InterHand2.6M/annotations'
        if self.mode == 'val':
            self.rootnet_output_path = '../data/InterHand2.6M/rootnet_output/rootnet_interhand2.6m_output_val.json'
        else:
            self.rootnet_output_path = '../data/InterHand2.6M/rootnet_output/rootnet_interhand2.6m_output_test.json'
        self.transform = transform
        self.joint_num = 21 # single hand
        self.root_joint_idx = {'right': 20, 'left': 41}
        self.joint_type = {'right': np.arange(0,self.joint_num), 'left': np.arange(self.joint_num,self.joint_num*2)}
        self.skeleton = load_skeleton(osp.join(self.annot_path, 'skeleton.txt'), self.joint_num*2)
        
        self.datalist = []
        self.datalist_sh = []
        self.datalist_ih = []
        self.sequence_names = []
        
        # load annotation
        print("Load annotation from  " + osp.join(self.annot_path, self.mode))
        db = COCO(osp.join(self.annot_path, self.mode, 'InterHand2.6M_' + self.mode + '_data.json'))
        with open(osp.join(self.annot_path, self.mode, 'InterHand2.6M_' + self.mode + '_camera.json')) as f:
            cameras = json.load(f)
        with open(osp.join(self.annot_path, self.mode, 'InterHand2.6M_' + self.mode + '_joint_3d.json')) as f:
            joints = json.load(f)

        if (self.mode == 'val' or self.mode == 'test') and cfg.trans_test == 'rootnet':
            print("Get bbox and root depth from " + self.rootnet_output_path)
            rootnet_result = {}
            with open(self.rootnet_output_path) as f:
                annot = json.load(f)
            for i in range(len(annot)):
                rootnet_result[str(annot[i]['annot_id'])] = annot[i]
        else:
            print("Get bbox and root depth from groundtruth annotation")
        
        for aid in db.anns.keys():
            ann = db.anns[aid]
            image_id = ann['image_id']
            img = db.loadImgs(image_id)[0]
 
            capture_id = img['capture']
            seq_name = img['seq_name']
            cam = img['camera']
            frame_idx = img['frame_idx']
            img_path = osp.join(self.img_path, self.mode, img['file_name'])
            
            campos, camrot = np.array(cameras[str(capture_id)]['campos'][str(cam)], dtype=np.float32), np.array(cameras[str(capture_id)]['camrot'][str(cam)], dtype=np.float32)
            focal, princpt = np.array(cameras[str(capture_id)]['focal'][str(cam)], dtype=np.float32), np.array(cameras[str(capture_id)]['princpt'][str(cam)], dtype=np.float32)
            joint_world = np.array(joints[str(capture_id)][str(frame_idx)]['world_coord'], dtype=np.float32)
            joint_cam = world2cam(joint_world.transpose(1,0), camrot, campos.reshape(3,1)).transpose(1,0)
            joint_img = cam2pixel(joint_cam, focal, princpt)[:,:2]

            joint_valid = np.array(ann['joint_valid'],dtype=np.float32).reshape(self.joint_num*2)
            # if root is not valid -> root-relative 3D pose is also not valid. Therefore, mark all joints as invalid
            joint_valid[self.joint_type['right']] *= joint_valid[self.root_joint_idx['right']]
            joint_valid[self.joint_type['left']] *= joint_valid[self.root_joint_idx['left']]
            hand_type = ann['hand_type']
            hand_type_valid = np.array((ann['hand_type_valid']), dtype=np.float32)
            
            if (self.mode == 'val' or self.mode == 'test') and cfg.trans_test == 'rootnet':
                bbox = np.array(rootnet_result[str(aid)]['bbox'],dtype=np.float32)
                abs_depth = {'right': rootnet_result[str(aid)]['abs_depth'][0], 'left': rootnet_result[str(aid)]['abs_depth'][1]}
            else:
                img_width, img_height = img['width'], img['height']
                bbox = np.array(ann['bbox'],dtype=np.float32) # x,y,w,h
                bbox = process_bbox(bbox, (img_height, img_width))
                abs_depth = {'right': joint_cam[self.root_joint_idx['right'],2], 'left': joint_cam[self.root_joint_idx['left'],2]}

            cam_param = {'focal': focal, 'princpt': princpt}
            joint = {'cam_coord': joint_cam, 'img_coord': joint_img, 'valid': joint_valid}
            data = {'img_path': img_path, 'seq_name': seq_name, 'cam_param': cam_param, 'bbox': bbox, 'joint': joint, 'hand_type': hand_type, 'hand_type_valid': hand_type_valid, 'abs_depth': abs_depth, 'file_name': img['file_name'], 'capture': capture_id, 'cam': cam, 'frame': frame_idx}
            if hand_type == 'right' or hand_type == 'left':
                self.datalist_sh.append(data)
            else:
                self.datalist_ih.append(data)
            if seq_name not in self.sequence_names:
                self.sequence_names.append(seq_name)

        self.datalist = self.datalist_sh + self.datalist_ih
        print('Number of annotations in single hand sequences: ' + str(len(self.datalist_sh)))
        print('Number of annotations in interacting hand sequences: ' + str(len(self.datalist_ih)))