Table · 表格#table#heat-cell#plot-from-image#wide-aspect
Class-wise IoU · 2-row heat-shaded comparison table
Class-wise IoU · 双行热度色块对比表
Long flat table (AR≈8:1) comparing Base vs Ours across 19 semantic-segmentation classes plus average. Each cell is a rectangle whose red intensity scales with the row's win margin against the other; winner cells with the largest gap flip to white text. Demonstrates the `plot-from-image` skill: an arbitrary paper figure rebuilt with raw matplotlib Rectangles + ax.text — no plt.table involved.
AR≈8:1 的长扁表格，对比 Base 与 Ours 在 19 个语义分割类别 + 平均上的 IoU。每个单元格是一个矩形：红色饱和度按该行相对另一行的胜出幅度递增，且大幅胜出格子的字体翻成白色。这一条目示范的是 `plot-from-image` 技能——用裸 matplotlib Rectangle + ax.text 直接搭表，而不是 plt.table。
@paper · 来自论文
Plot-from-image · Class-wise IoU demo (community paper figure)
plot-from-image · Class-wise IoU 演示（用户上传论文截图）
paper_experiment_plot_skills contributors · community 2025
// original from paper · 论文原图
// reproduced via classwise_iou.py · 脚本复现download png
classwise_iou.py
"""
Reproduce: Class-Wise IoU Results
Original: 1671x209 px AR=8.00
"""
from __future__ import annotations
import matplotlib.pyplot as plt
import numpy as np
from matplotlib.patches import Rectangle
LABELS = [
"road",
"build.",
"sky",
"side.",
"veget.",
"car",
"terr.",
"wall",
"truck",
"pole",
"fence",
"bus",
"pers.",
"light",
"sign.",
"train",
"moto.",
"rider",
"bicy.",
"avg.",
]
BASE = np.array(
[
93.7,
90.5,
91.6,
63.3,
90.5,
90.8,
45.3,
52.2,
45.0,
53.3,
47.5,
76.8,
78.5,
66.9,
49.1,
55.6,
55.7,
43.5,
54.8,
65.5,
]
)
OURS = np.array(
[
93.3,
90.9,
90.3,
59.0,
90.1,
92.2,
43.1,
53.2,
54.1,
56.6,
50.4,
80.9,
80.8,
66.8,
51.2,
70.6,
63.4,
56.2,
65.7,
68.9,
]
)
def blend(color_a: str, color_b: str, ratio: float) -> tuple[float, float, float]:
"""Blend two hex colors."""
ratio = float(np.clip(ratio, 0.0, 1.0))
a = np.array([int(color_a[i : i + 2], 16) for i in (1, 3, 5)], dtype=float)
b = np.array([int(color_b[i : i + 2], 16) for i in (1, 3, 5)], dtype=float)
return tuple(((1 - ratio) * a + ratio * b) / 255.0)
def cell_color(delta_strength: float, is_winner: bool) -> tuple[float, float, float]:
"""
Approximate the paper's heat shading:
winner row gets stronger red, loser row stays very light.
"""
if is_winner:
ratio = 0.10 + 0.82 * delta_strength
else:
ratio = 0.02 + 0.20 * delta_strength
return blend("#ffffff", "#c81419", ratio)
def text_color(delta_strength: float, is_winner: bool) -> str:
if is_winner and delta_strength >= 0.72:
return "white"
return "#1b1b1b"
def draw_row(
ax: plt.Axes,
values: np.ndarray,
other_values: np.ndarray,
row_label: str,
y0: float,
cell_w: float,
cell_h: float,
gap: float,
x0: float,
) -> None:
deltas = values - other_values
strengths = np.clip(np.abs(deltas) / 14.0, 0.0, 1.0)
ax.text(
x0 - 0.22,
y0 + cell_h / 2,
row_label,
ha="right",
va="center",
fontsize=13.8,
family="sans-serif",
color="#111111",
)
for i, val in enumerate(values):
is_winner = val >= other_values[i]
x = x0 + i * (cell_w + gap)
face = cell_color(strengths[i], is_winner)
ax.add_patch(
Rectangle(
(x, y0),
cell_w,
cell_h,
facecolor=face,
edgecolor="none",
)
)
ax.text(
x + cell_w / 2,
y0 + cell_h / 2,
f"{val:.1f}",
ha="center",
va="center",
fontsize=12.1,
family="sans-serif",
color=text_color(strengths[i], is_winner),
)
def main() -> None:
plt.rcParams.update(
{
"font.family": "sans-serif",
"font.sans-serif": ["Arial", "Helvetica", "DejaVu Sans"],
}
)
n_cols = len(LABELS)
cell_w = 0.96
cell_h = 0.42
gap = 0.035
x0 = 0.98
total_w = x0 + n_cols * (cell_w + gap) - gap + 0.10
fig, ax = plt.subplots(figsize=(16.0, 1.78))
ax.set_xlim(0, total_w)
ax.set_ylim(-0.58, 1.48)
ax.axis("off")
draw_row(ax, BASE, OURS, "Base", y0=0.86, cell_w=cell_w, cell_h=cell_h, gap=gap, x0=x0)
draw_row(ax, OURS, BASE, "Ours", y0=0.38, cell_w=cell_w, cell_h=cell_h, gap=gap, x0=x0)
for i, label in enumerate(LABELS):
x = x0 + i * (cell_w + gap) + cell_w / 2
ax.text(
x,
0.11,
label,
ha="center",
va="center",
fontsize=11.8,
family="sans-serif",
color="#111111",
)
ax.text(
total_w / 2,
-0.30,
"(b) Class-Wise IoU Results",
ha="center",
va="center",
fontsize=15.4,
family="serif",
color="#111111",
)
fig.savefig(
"classwise_iou_repro.png",
dpi=300,
facecolor="white",
bbox_inches="tight",
pad_inches=0.02,
)
plt.close(fig)
if __name__ == "__main__":
main()