|
| 1 | +# Copyright (C) 2022 CVAT.ai Corporation |
| 2 | +# |
| 3 | +# SPDX-License-Identifier: MIT |
| 4 | + |
| 5 | +import jsonpickle |
| 6 | +import numpy as np |
| 7 | +import torch |
| 8 | +from pysot_toolkit.bbox import get_axis_aligned_bbox |
| 9 | +from pysot_toolkit.trackers.net_wrappers import NetWithBackbone |
| 10 | +from pysot_toolkit.trackers.tracker import Tracker |
| 11 | + |
| 12 | + |
| 13 | +class ModelHandler: |
| 14 | + def __init__(self): |
| 15 | + use_gpu = torch.cuda.is_available() |
| 16 | + net_path = '/transt.pth' # Absolute path of the model |
| 17 | + net = NetWithBackbone(net_path=net_path, use_gpu=use_gpu) |
| 18 | + self.tracker = Tracker(name='transt', net=net, window_penalty=0.49, exemplar_size=128, instance_size=256) |
| 19 | + |
| 20 | + def decode_state(self, state): |
| 21 | + self.tracker.net.net.zf = jsonpickle.decode(state['model.net.net.zf']) |
| 22 | + self.tracker.net.net.pos_template = jsonpickle.decode(state['model.net.net.pos_template']) |
| 23 | + |
| 24 | + self.tracker.window = jsonpickle.decode(state['model.window']) |
| 25 | + self.tracker.center_pos = jsonpickle.decode(state['model.center_pos']) |
| 26 | + self.tracker.size = jsonpickle.decode(state['model.size']) |
| 27 | + self.tracker.channel_average = jsonpickle.decode(state['model.channel_average']) |
| 28 | + self.tracker.mean = jsonpickle.decode(state['model.mean']) |
| 29 | + self.tracker.std = jsonpickle.decode(state['model.std']) |
| 30 | + self.tracker.inplace = jsonpickle.decode(state['model.inplace']) |
| 31 | + |
| 32 | + self.tracker.features_initialized = False |
| 33 | + if 'model.features_initialized' in state: |
| 34 | + self.tracker.features_initialized = jsonpickle.decode(state['model.features_initialized']) |
| 35 | + |
| 36 | + def encode_state(self): |
| 37 | + state = {} |
| 38 | + state['model.net.net.zf'] = jsonpickle.encode(self.tracker.net.net.zf) |
| 39 | + state['model.net.net.pos_template'] = jsonpickle.encode(self.tracker.net.net.pos_template) |
| 40 | + state['model.window'] = jsonpickle.encode(self.tracker.window) |
| 41 | + state['model.center_pos'] = jsonpickle.encode(self.tracker.center_pos) |
| 42 | + state['model.size'] = jsonpickle.encode(self.tracker.size) |
| 43 | + state['model.channel_average'] = jsonpickle.encode(self.tracker.channel_average) |
| 44 | + state['model.mean'] = jsonpickle.encode(self.tracker.mean) |
| 45 | + state['model.std'] = jsonpickle.encode(self.tracker.std) |
| 46 | + state['model.inplace'] = jsonpickle.encode(self.tracker.inplace) |
| 47 | + state['model.features_initialized'] = jsonpickle.encode(getattr(self.tracker, 'features_initialized', False)) |
| 48 | + |
| 49 | + return state |
| 50 | + |
| 51 | + def init_tracker(self, img, bbox): |
| 52 | + cx, cy, w, h = get_axis_aligned_bbox(np.array(bbox)) |
| 53 | + gt_bbox_ = [cx - w / 2, cy - h / 2, w, h] |
| 54 | + init_info = {'init_bbox': gt_bbox_} |
| 55 | + self.tracker.initialize(img, init_info) |
| 56 | + |
| 57 | + def track(self, img): |
| 58 | + outputs = self.tracker.track(img) |
| 59 | + prediction_bbox = outputs['target_bbox'] |
| 60 | + |
| 61 | + left = prediction_bbox[0] |
| 62 | + top = prediction_bbox[1] |
| 63 | + right = prediction_bbox[0] + prediction_bbox[2] |
| 64 | + bottom = prediction_bbox[1] + prediction_bbox[3] |
| 65 | + return (left, top, right, bottom) |
| 66 | + |
| 67 | + def infer(self, image, shape, state): |
| 68 | + if state is None: |
| 69 | + init_shape = (shape[0], shape[1], shape[2] - shape[0], shape[3] - shape[1]) |
| 70 | + |
| 71 | + self.init_tracker(image, init_shape) |
| 72 | + state = self.encode_state() |
| 73 | + else: |
| 74 | + self.decode_state(state) |
| 75 | + shape = self.track(image) |
| 76 | + state = self.encode_state() |
| 77 | + |
| 78 | + return shape, state |
0 commit comments