File size: 5,012 Bytes
4d4dd90
8320ccc
4d4dd90
8320ccc
 
 
 
4d4dd90
8320ccc
4d4dd90
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
8320ccc
4d4dd90
8320ccc
4d4dd90
 
 
 
 
 
 
 
 
 
 
 
 
 
8320ccc
4d4dd90
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import subprocess
import sys
from pathlib import Path

import gdown
import torch

from .. import logger
from ..utils.base_model import BaseModel

gim_path = Path(__file__).parent / "../../third_party/gim"
sys.path.append(str(gim_path))

from dkm.models.model_zoo.DKMv3 import DKMv3

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")


class GIM(BaseModel):
    default_conf = {
        "model_name": "gim_dkm_100h.ckpt",
        "match_threshold": 0.2,
        "checkpoint_dir": gim_path / "weights",
    }
    required_inputs = [
        "image0",
        "image1",
    ]
    model_dict = {
        "gim_lightglue_100h.ckpt": "https://github.com/xuelunshen/gim/blob/main/weights/gim_lightglue_100h.ckpt",
        "gim_dkm_100h.ckpt": "https://drive.google.com/file/d/1gk97V4IROnR1Nprq10W9NCFUv2mxXR_-/view",
    }

    def _init(self, conf):
        conf["model_name"] = str(conf["weights"])
        if conf["model_name"] not in self.model_dict:
            raise ValueError(f"Unknown GIM model {conf['model_name']}.")
        model_path = conf["checkpoint_dir"] / conf["model_name"]

        # Download the model.
        if not model_path.exists():
            model_path.parent.mkdir(exist_ok=True)
            model_link = self.model_dict[conf["model_name"]]
            if "drive.google.com" in model_link:
                gdown.download(model_link, output=str(model_path), fuzzy=True)
            else:
                cmd = ["wget", "--quiet", model_link, "-O", str(model_path)]
                subprocess.run(cmd, check=True)
            logger.info("Downloaded GIM model succeeed!")

        self.aspect_ratio = 896 / 672
        model = DKMv3(None, 672, 896, upsample_preds=True)
        state_dict = torch.load(str(model_path), map_location="cpu")
        if "state_dict" in state_dict.keys():
            state_dict = state_dict["state_dict"]
        for k in list(state_dict.keys()):
            if k.startswith("model."):
                state_dict[k.replace("model.", "", 1)] = state_dict.pop(k)
            if "encoder.net.fc" in k:
                state_dict.pop(k)
        model.load_state_dict(state_dict)

        self.net = model
        logger.info("Loaded GIM model")

    def pad_image(self, image, aspect_ratio):
        new_width = max(image.shape[3], int(image.shape[2] * aspect_ratio))
        new_height = max(image.shape[2], int(image.shape[3] / aspect_ratio))
        pad_width = new_width - image.shape[3]
        pad_height = new_height - image.shape[2]
        return torch.nn.functional.pad(
            image,
            (
                pad_width // 2,
                pad_width - pad_width // 2,
                pad_height // 2,
                pad_height - pad_height // 2,
            ),
        )

    def rescale_kpts(self, sparse_matches, shape0, shape1):
        kpts0 = torch.stack(
            (
                shape0[1] * (sparse_matches[:, 0] + 1) / 2,
                shape0[0] * (sparse_matches[:, 1] + 1) / 2,
            ),
            dim=-1,
        )
        kpts1 = torch.stack(
            (
                shape1[1] * (sparse_matches[:, 2] + 1) / 2,
                shape1[0] * (sparse_matches[:, 3] + 1) / 2,
            ),
            dim=-1,
        )
        return kpts0, kpts1

    def compute_mask(self, kpts0, kpts1, orig_shape0, orig_shape1):
        mask = (
            (kpts0[:, 0] > 0)
            & (kpts0[:, 1] > 0)
            & (kpts1[:, 0] > 0)
            & (kpts1[:, 1] > 0)
        )
        mask &= (
            (kpts0[:, 0] <= (orig_shape0[1] - 1))
            & (kpts1[:, 0] <= (orig_shape1[1] - 1))
            & (kpts0[:, 1] <= (orig_shape0[0] - 1))
            & (kpts1[:, 1] <= (orig_shape1[0] - 1))
        )
        return mask

    def _forward(self, data):
        image0, image1 = self.pad_image(
            data["image0"], self.aspect_ratio
        ), self.pad_image(data["image1"], self.aspect_ratio)
        dense_matches, dense_certainty = self.net.match(image0, image1)
        sparse_matches, mconf = self.net.sample(
            dense_matches, dense_certainty, self.conf["max_keypoints"]
        )
        kpts0, kpts1 = self.rescale_kpts(
            sparse_matches, image0.shape[-2:], image1.shape[-2:]
        )
        mask = self.compute_mask(
            kpts0, kpts1, data["image0"].shape[-2:], data["image1"].shape[-2:]
        )
        b_ids, i_ids = torch.where(mconf[None])
        pred = {
            "keypoints0": kpts0[i_ids],
            "keypoints1": kpts1[i_ids],
            "confidence": mconf[i_ids],
            "batch_indexes": b_ids,
        }
        scores, b_ids = pred["confidence"], pred["batch_indexes"]
        kpts0, kpts1 = pred["keypoints0"], pred["keypoints1"]
        pred["confidence"], pred["batch_indexes"] = scores[mask], b_ids[mask]
        pred["keypoints0"], pred["keypoints1"] = kpts0[mask], kpts1[mask]

        out = {
            "keypoints0": pred["keypoints0"],
            "keypoints1": pred["keypoints1"],
        }
        return out