File size: 7,771 Bytes
9882c88
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
from math import pi, sqrt

import matplotlib.pyplot as plt
import torch
import torch.nn.functional as F

import anycalib.visualization.viz_2d as viz_2d
from anycalib.cameras.base import BaseCamera
from anycalib.manifolds import Unit3

RAD2DEG = 180 / pi


def create_figures(im_size, nrows, ncols, min_fig_size: float = 4.0):
    h, w = im_size
    plot_w = min_fig_size if w < h else min_fig_size * w / h
    plot_h = min_fig_size if h < w else min_fig_size * h / w
    fig, axs = plt.subplots(
        nrows,
        ncols,
        figsize=(plot_w * ncols, plot_h * nrows),
        squeeze=False,
    )
    return fig, axs


def make_radial_figure(
    pred: dict,
    data: dict,
    n_pairs: int = 2,
    min_fig_size: float = 4.0,
    show_radial_vecs: bool = False,
) -> dict:
    """Image grid of radial vectors plots."""
    titles = ("Image", "Polar angles GT", "Pred")
    if show_radial_vecs:
        titles = titles + ("Radial vecs GT", "Pred")
    if "tangent_coords" in pred:
        titles = titles + ("Tangent θx GT", "Pred", "Tangent θy GT", "Pred")

    # image grid
    n_pairs = min(n_pairs, len(data["image"]))
    b, _, h, w = data["image"].shape
    fig, axs = create_figures((h, w), n_pairs, len(titles), min_fig_size=min_fig_size)

    images = data["image"].permute(0, 2, 3, 1).clamp(0, 1)
    rays_gt = data["rays"].view(b, h, w, 3)
    rays = pred["rays"].view(b, h, w, 3)
    if "tangent_coords" in pred:
        tcoords_gt = (RAD2DEG * Unit3.logmap_at_z1(rays_gt)).cpu()
        tcoords = (RAD2DEG * pred["tangent_coords"].detach().view(b, h, w, 2)).cpu()
        tc_kwargs = {
            "contours_every": 10,
            "cmap": "Spectral_r",
            "vmin": -90,
            "vmax": 90,
        }

    for i in range(n_pairs):
        axr = axs[i]
        for j, title in enumerate(titles):
            viz_2d.plot_image(axr[j], images[i])
            if i == 0:
                axr[j].set_title(title)
        # polar angles
        viz_2d.plot_polar_angles(axr[1], rays_gt[i])
        viz_2d.plot_polar_angles(axr[2], rays[i])
        j = 3
        if show_radial_vecs:
            # quiver plots with the same scale as the ground-truth
            viz_2d.plot_radial_vectors(axr[j], rays_gt[i], color_vectors="lime")
            fig.canvas.draw()  # ensure the scale attribute of the quiver plot is present
            viz_2d.plot_radial_vectors(
                axr[j + 1],
                rays[i],
                color_vectors="orange",
                scale=axr[1].collections[-1].scale,
            )
            j = 5
        if "tangent_coords" in pred:
            tc_gt = tcoords_gt[i]
            tc = tcoords[i]
            viz_2d.plot_contours(axr[j], tc_gt[..., 0], **tc_kwargs)
            viz_2d.plot_contours(axr[j + 1], tc[..., 0], **tc_kwargs)
            viz_2d.plot_contours(axr[j + 2], tc_gt[..., 1], **tc_kwargs)
            viz_2d.plot_contours(axr[j + 3], tc[..., 1], **tc_kwargs)
    fig.tight_layout()
    return {"radial": fig}


def make_errors_figure(
    pred: dict,
    data: dict,
    n_pairs: int = 2,
    min_fig_size: float = 4.0,
    log_normalize: bool = False,
    agg_fn: str | None = None,  # "prod",
) -> dict:
    """Image grid of error plots."""
    titles = ("Image", "Tangent Error", "Angular Error")
    suptitle = "Errors"
    if "log_covs" in pred:
        suptitle = "Errors and Uncertainties"
        unc_str = "log-uncertainty" if log_normalize else "uncertainty"
        titles += (f"{unc_str}-x", f"{unc_str}-y") if agg_fn is None else (unc_str,)
    if "weights" in pred:
        titles += ("alpha",)

    # image grid
    n_pairs = min(n_pairs, len(data["image"]))
    b, _, h, w = data["image"].shape
    fig, axs = create_figures((h, w), n_pairs, len(titles), min_fig_size=min_fig_size)
    fig.suptitle(suptitle)

    images = data["image"].permute(0, 2, 3, 1).clamp(0, 1)
    rays_gt = data["rays"].view(b, h, w, 3)
    rays = pred["rays"].view(b, h, w, 3)
    log_covs = pred.get("log_covs", None)
    if log_covs is not None:
        covs = log_covs.exp().view(b, h, w, 2)
    if "weights" in pred:
        # compute "alpha(s)" of mixture model
        alpha = torch.softmax(pred["weights"], dim=-1)[..., 0]  # (b, 1 or h*w)
        alpha = (
            alpha.expand(-1, h * w).view(b, h, w).cpu()
            if alpha.shape[-1] == 1
            else alpha.view(b, h, w).cpu()
        )

    for i in range(n_pairs):
        axr = axs[i]
        for j, title in enumerate(titles):
            viz_2d.plot_image(axr[j], images[i])
            if i == 0:
                axr[j].set_title(title)
        # errors
        viz_2d.plot_tangent_errors_as_vectors(axr[1], rays[i], rays_gt[i])
        viz_2d.plot_angular_errors(axr[2], rays[i], rays_gt[i], add_colorbar=True)
        # uncertainties
        if log_covs is not None:
            axr_u = axr[3] if isinstance(agg_fn, str) else (axr[3], axr[4])
            viz_2d.plot_uncertainties_as_heatmap(
                axr_u,
                covs[i],
                log_normalize=log_normalize,
                aggregator=agg_fn,
                add_colorbar=True,
            )
        if "weights" in pred:
            viz_2d.plot_heatmap(axr[-1], alpha[i], add_colorbar=True)

    fig.tight_layout()
    return {"errors": fig}


def make_editmaps_figure(
    pred: dict, data: dict, n_pairs: int = 2, min_fig_size: float = 4.0
):
    """Image grid of editmap plots."""
    titles = ("Image", "Pix AR Error", "Uncertainty", "Radii GT [pix]", "Pred")
    n_pairs = min(n_pairs, len(data["image"]))
    _, _, h, w = data["image"].shape
    rmax = 0.5 * sqrt(h**2 + w**2)
    fig, axs = create_figures((h, w), n_pairs, len(titles), min_fig_size=min_fig_size)

    images = data["image"].permute(0, 2, 3, 1).clamp(0, 1)
    _, _, hp, wp = pred["pix_ar_map"].shape
    assert h / hp == w / wp and h / hp >= 1

    pix_ar_gt = data["pix_ar"]  # (b,)
    radii = (h / hp) * pred["radii"].detach()  # (b, hp, wp)
    # upsample to image resolution
    radii = F.interpolate(radii[:, None], (h, w), mode="bilinear", align_corners=False)[:, 0].cpu()  # fmt:skip
    pix_ar = F.interpolate(
        pred["pix_ar_map"].detach(), (h, w), mode="bilinear", align_corners=False
    )  # (b, 2, h, w)

    pix_ar_err = (pix_ar[:, 0] - pix_ar_gt[:, None, None]).abs().cpu()  # (b, h, w)
    pix_ar_unc = torch.exp(pix_ar[:, 1]).cpu()
    radii_gt = torch.linalg.norm(
        BaseCamera.pixel_grid_coords(h, w, data["cxcy_gt"], 0.5)
        - data["cxcy_gt"][:, None, None],
        dim=-1,
    ).cpu()

    for i in range(n_pairs):
        axr = axs[i]
        for j, title in enumerate(titles):
            viz_2d.plot_image(axr[j], images[i])
            if i == 0:
                axr[j].set_title(title)
        viz_2d.plot_heatmap(axr[1], pix_ar_err[i], alpha=0.5, add_colorbar=True, cmap="error")  # fmt: skip
        viz_2d.plot_text(axr[1], f"GT: {pix_ar_gt[i].item():.1f}")
        viz_2d.plot_heatmap(axr[2], pix_ar_unc[i], alpha=0.5, add_colorbar=True, cmap="turbo_r")  # fmt: skip
        viz_2d.plot_contours(axr[3], radii_gt[i], vmin=0, vmax=rmax, contours_every=25, label_units="")  # fmt: skip
        viz_2d.plot_contours(axr[4], radii[i], vmin=0, vmax=rmax, contours_every=25, label_units="")  # fmt: skip

    fig.tight_layout()
    return {"editmaps": fig}


def make_batch_figures(
    pred: dict, data: dict, n_pairs: int = 3, min_fig_size: float = 3.0
) -> dict:
    """Create figures for debugging"""
    figs = make_radial_figure(pred, data, n_pairs, min_fig_size=min_fig_size)
    figs |= make_errors_figure(pred, data, n_pairs, min_fig_size=min_fig_size)
    if "pix_ar_map" in pred and "radii" in pred:
        figs |= make_editmaps_figure(pred, data, n_pairs, min_fig_size=min_fig_size)
    return figs