POPE/eval_ycb_json.py at main · karltan0328/POPE · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
from pope_model_api import *

if __name__  == "__main__":
    ckpt, model_type = get_model_info("h")
    sam = sam_model_registry[model_type](checkpoint=ckpt)
    DEVICE = "cuda"
    sam.to(device=DEVICE)
    MASK_GEN = SamAutomaticMaskGenerator(sam)
    logger.info(f"load SAM model from {ckpt}")
    crop_tool = CropImage()
    dinov2_model = load_dinov2_model()
    dinov2_model.to("cuda:0")

    metrics = dict()
    metrics.update({'R_errs':[], 't_errs':[], 'inliers':[], "identifiers":[]})

    ROOT_DIR = "data/ycbv/"
    res_table = []

    import json
    with open("data/pairs/YCB-VIDEO-test.json") as f:
        dir_list = json.load(f)

    for label_idx , test_dict in enumerate(dir_list):
        logger.info(f"YCB-VIDEO: {label_idx}")
        metrics = dict()
        metrics.update({'R_errs':[], 't_errs':[], 'inliers':[] , "identifiers":[]})
        sample_data = dir_list[label_idx]["0"][0]
        label = sample_data.split("/")[0]
        name = label.split("-")[1]
        dir_name = os.path.dirname(sample_data)
        FULL_ROOT_DIR = os.path.join(ROOT_DIR, dir_name)
        recall_image,all_image = 0,0
        for rotation_key, rotation_list in zip(test_dict.keys(), test_dict.values()):
            for pair_idx,pair_name in enumerate(tqdm(rotation_list[::2])):
                all_image = all_image + 1
                base_name = os.path.basename(pair_name)
                idx0_name = base_name.split("png-")[0]+"png"
                idx1_name = base_name.split("png-")[1]
                image0_name = os.path.join(FULL_ROOT_DIR, idx0_name)
                image1_name = os.path.join(FULL_ROOT_DIR.replace("color", "color_full"), idx1_name)
                intrinsic_path = image0_name.replace("color", "intrin_ba").replace("png","txt")
                K0 = np.loadtxt(intrinsic_path, delimiter=' ')
                intrinsic_path = image1_name.replace("color_full", "intrin").replace("png", "txt")
                K1 = np.loadtxt(intrinsic_path, delimiter=' ')
                image0 = cv2.imread(image0_name)
                ref_torch_image = set_torch_image(image0, center_crop=True)
                ref_fea = get_cls_token_torch(dinov2_model, ref_torch_image)
                image1 = cv2.imread(image1_name)
                image_h, image_w, _ = image1.shape
                t1 = time.time()
                masks = MASK_GEN.generate(image1)
                t2 = time.time()
                similarity_score, top_images = np.array([0, 0, 0], np.float32), [[], [], []]
                t3 = time.time()
                compact_percent = 0.3
                for xxx, mask in enumerate(masks):
                    object_mask = np.expand_dims(mask["segmentation"], -1)
                    x0, y0, w, h = mask["bbox"]
                    x1, y1 = x0 + w, y0 + h
                    x0 -= int(w * compact_percent)
                    y0 -= int(h * compact_percent)
                    x1 += int(w * compact_percent)
                    y1 += int(h * compact_percent)
                    box = np.array([x0, y0, x1, y1])
                    resize_shape = np.array([y1 - y0, x1 - x0])
                    K_crop, K_crop_homo = get_K_crop_resize(box, K1, resize_shape)
                    image_crop, _ = get_image_crop_resize(image1, box, resize_shape)
                    # object_mask, _ = get_image_crop_resize(object_mask, box, resize_shape)
                    box_new = np.array([0, 0, x1 - x0, y1 - y0])
                    resize_shape = np.array([256, 256])
                    K_crop, K_crop_homo = get_K_crop_resize(box_new, K_crop, resize_shape)
                    image_crop, _ = get_image_crop_resize(image_crop, box_new, resize_shape)
                    crop_tensor = set_torch_image(image_crop, center_crop=True)
                    with torch.no_grad():
                        fea = get_cls_token_torch(dinov2_model, crop_tensor)
                    score = F.cosine_similarity(ref_fea, fea, dim=1, eps=1e-8)
                    if  (score.item() > similarity_score).any():
                        mask["crop_image"] = image_crop
                        mask["K"] = K_crop
                        mask["bbox"] = box
                        min_idx = np.argmin(similarity_score)
                        similarity_score[min_idx] = score.item()
                        top_images[min_idx] = mask.copy()

                img0 =  cv2.cvtColor(image0, cv2.COLOR_BGR2GRAY)
                img0 = torch.from_numpy(img0).float()[None] / 255.
                img0 = img0.unsqueeze(0).cuda()

                matching_score = [[0] for _ in range(len(top_images))]
                for top_idx in range(len(top_images)):
                    img1 =  cv2.cvtColor(top_images[top_idx]["crop_image"], cv2.COLOR_BGR2GRAY)
                    img1 = torch.from_numpy(img1).float()[None] / 255.
                    img1 = img1.unsqueeze(0).cuda()
                    batch = {'image0':img0, 'image1':img1}
                    with torch.no_grad():
                        matcher(batch)
                        mkpts0 = batch['mkpts0_f'].cpu().numpy()
                        mkpts1 = batch['mkpts1_f'].cpu().numpy()
                        confidences = batch["mconf"].cpu().numpy()
                    conf_mask = np.where(confidences > 0.9)
                    matching_score[top_idx] = conf_mask[0].shape[0]
                    top_images[top_idx]["mkpts0"] = mkpts0
                    top_images[top_idx]["mkpts1"] = mkpts1
                    top_images[top_idx]["mconf"] = confidences
                #---------------------------------------------------
                # crop_image = cv2.resize(top_images[np.argmax(matching_score)]["crop_image"], (256, 256))
                # que_image = cv2.resize(image0, (256, 256))
                # image = np.hstack((que_image, crop_image))
                # for top_idx in range(len(top_images)):
                #     crop_image = top_images[top_idx]["crop_image"]
                #     score = matching_score[top_idx]
                #     crop_image = cv2.resize(crop_image, (256, 256))
                #     cv2.putText(crop_image,f'{score}', (100, 100), cv2.FONT_HERSHEY_COMPLEX, 1, (0, 0, 255), 1)
                #     image = np.hstack((image, crop_image))
                # cv2.imwrite(f"segment_anything/crop_images/{idx}.jpg", image)
                #---------------------------------------------------
                t4 = time.time()
                # print(f"t4-t3: object detection:{1000*(t4-t3)} ms")
                pose0_name = image0_name.replace("color", "poses_ba").replace("png", "txt")
                pose1_name = image1_name.replace("color_full", "poses_ba").replace("png", "txt")
                pose0 = np.loadtxt(pose0_name)
                pose1 = np.loadtxt(pose1_name)
                pose0 = np.vstack((pose0, np.array([[0, 0, 0, 1]])))
                pose1 = np.vstack((pose1, np.array([[0, 0, 0, 1]])))
                relative_pose = np.matmul(pose1, inv(pose0))
                t = relative_pose[:3, -1].reshape(1, 3)

                max_match_idx = np.argmax(matching_score)
                pre_bbox = top_images[max_match_idx]["bbox"]
                mkpts0 = top_images[max_match_idx]["mkpts0"]
                mkpts1 = top_images[max_match_idx]["mkpts1"]
                pre_K = top_images[max_match_idx]["K"]

                gt_bbox_name = image0_name.replace("color", "bbox_2d").replace("png", "txt")

                gt_bbox = np.loadtxt(gt_bbox_name)
                is_recalled = recall_object(pre_bbox, gt_bbox)
                recall_image = recall_image + int(is_recalled > 0.5)
                ret = estimate_pose(mkpts0, mkpts1, K0, pre_K, 0.5, 0.99)
                if ret is not None:
                    Rot, t, inliers = ret
                    t_err, R_err = relative_pose_error(relative_pose, Rot, t, ignore_gt_t_thr=0.0)
                    metrics['R_errs'].append(R_err)
                    metrics['t_errs'].append(t_err)
                else:
                    metrics['R_errs'].append(90)
                    metrics['t_errs'].append(90)
                metrics["identifiers"].append(pair_name)

        import pprint
        from src.utils.metrics import (
            aggregate_metrics
        )
        from loguru import logger
        val_metrics_4tb = aggregate_metrics(metrics, 5e-4)
        val_metrics_4tb["AP50"] = recall_image / all_image
        logger.info('\n' + pprint.pformat(val_metrics_4tb))
        res_table.append( [f"{name}"] + list(val_metrics_4tb.values()))

    from tabulate import tabulate
    headers = ["Category"] + list(val_metrics_4tb.keys())
    all_data = np.array(res_table)[:,1:].astype(np.float32)
    res_table.append(["Avg"] + all_data.mean(0).tolist())
    print(tabulate(res_table, \
        headers=headers, tablefmt='fancy_grid'))