import matplotlib.pyplot as plt import matplotlib.patches as mpatches import matplotlib.lines as mlines import numpy as np from adjustText import adjust_text # ── 1. DATA ─────────────────────────────────────────────────────────────────── data = [ dict(name='Mnist\n(2015)', mae=7.29, gflops=0.10, params=1.82, group='prior'), dict(name='iTracker\n(2016)', mae=7.67, gflops=3.97, params=6.28, group='prior'), dict(name='GazeNet\n(2017)', mae=6.62, gflops=72.24, params=90.23, group='prior'), dict(name='FullFace\n(2017)', mae=5.65, gflops=29.90, params=190.00, group='prior'), dict(name='RT-Gene\n(2018)', mae=5.36, gflops=12.21, params=31.66, group='prior'), dict(name='DilatedNet\n(2019)', mae=5.07, gflops=202.00, params=3.92, group='prior'), dict(name='Gaze360\n(2019)', mae=4.66, gflops=3.65, params=11.72, group='recent'), dict(name='FAR-Net\n(2021)', mae=5.12, gflops=0.65, params=1.94, group='recent'), dict(name='FR-Net\n(2024)', mae=4.95, gflops=0.15, params=0.85, group='recent'), dict(name='FGI-Net\n(2025)', mae=4.81, gflops=0.08, params=0.45, group='recent'), dict(name='Heavy Teacher\n(ResNet50)', mae=4.15, gflops=4.12, params=25.56, group='teacher'), dict(name='LIPE \n(Ours)', mae=4.72, gflops=0.02125, params=0.18, group='ours'), ] # ── 2. STYLE ────────────────────────────────────────────────────────────────── plt.rcParams['font.family'] = 'DejaVu Sans' plt.rcParams['xtick.direction'] = 'in' plt.rcParams['ytick.direction'] = 'in' GROUP_STYLE = { 'prior': {'color': '#C8C5BC', 'edgecolor': '#5F5E5A', 'label': 'Prior Works (2015–2019)'}, 'recent': {'color': '#6B8E23', 'edgecolor': '#3A5010', 'label': 'Recent Edge SOTA'}, 'teacher': {'color': '#9E9E9E', 'edgecolor': '#555555', 'label': 'Heavy Teacher Baseline'}, 'ours': {'color': '#E53935', 'edgecolor': '#7B1FA2', 'label': 'LIPE (Ours)'}, } ARROW_STYLE = dict( arrowstyle='->', color='#888888', lw=0.9, connectionstyle='arc3,rad=0.0' ) fig, ax = plt.subplots(figsize=(9, 6.5), dpi=200) fig.patch.set_facecolor('#FAFAFA') ax.set_facecolor('#FAFAFA') def bubble_area(p): return 90 + 380 * np.log10(p + 1) # ── 3. SCATTER POINTS ───────────────────────────────────────────────────────── for d in data: style = GROUP_STYLE[d['group']] s = bubble_area(d['params']) marker = 'D' if d['group'] == 'ours' else 'o' lw = 2.0 if d['group'] == 'ours' else 0.8 zo = 6 if d['group'] == 'ours' else (5 if d['group'] == 'teacher' else 3) ax.scatter( d['gflops'], d['mae'], s=s, color=style['color'], edgecolors=style['edgecolor'], marker=marker, alpha=0.88, linewidths=lw, zorder=zo, ) # ── 4. ANNOTATIONS WITH ARROWS (adjustText for collision avoidance) ─────────── # Manual offsets (in data-space offsets via display transform) — fine-tuned per label # key: point name (first line), value: (dx_pts, dy_pts) offset for text MANUAL_OFFSETS = { 'LIPE \n(Ours)': (-68, 32), 'FGI-Net\n(2025)': ( 12, -28), 'FR-Net\n(2024)': ( 12, 22), 'FAR-Net\n(2021)': (-70, -12), 'Gaze360\n(2019)': ( 14, 22), 'Heavy Teacher\n(ResNet50)': ( 14, 10), 'DilatedNet\n(2019)': ( 14, 0), 'RT-Gene\n(2018)': ( 14, 10), 'FullFace\n(2017)': ( 14, -26), 'GazeNet\n(2017)': ( 14, 0), 'iTracker\n(2016)': (-72, 10), 'Mnist\n(2015)': ( 14, -26), } texts = [] arrows = [] for d in data: is_ours = (d['group'] == 'ours') fw = 'bold' if is_ours else 'normal' col = '#C62828' if is_ours else '#333333' fs = 8.5 if is_ours else 7.5 bg_alpha = 0.82 if is_ours else 0.70 bg_color = '#FFF9F9' if is_ours else 'white' dx, dy = MANUAL_OFFSETS.get(d['name'], (12, 0)) ann = ax.annotate( d['name'], xy=(d['gflops'], d['mae']), xytext=(dx, dy), textcoords='offset points', fontsize=fs, fontweight=fw, color=col, ha='center', va='center', zorder=9, bbox=dict( boxstyle='round,pad=0.28', fc=bg_color, ec='#CCCCCC' if not is_ours else '#E57373', lw=0.6 if not is_ours else 1.0, alpha=bg_alpha, ), arrowprops=dict( arrowstyle='->', color='#AAAAAA' if not is_ours else '#E53935', lw=0.85 if not is_ours else 1.2, connectionstyle='arc3,rad=0.15', ), ) texts.append(ann) # ── 5. PARETO FRONTIER ──────────────────────────────────────────────────────── pareto_x = [0.02125, 0.08, 3.65, 4.12] pareto_y = [4.72, 4.81, 4.66, 4.15] ax.plot(pareto_x, pareto_y, color='#888888', linestyle='--', linewidth=1.3, zorder=2, alpha=0.8) # ── 6. AXES ─────────────────────────────────────────────────────────────────── ax.set_xscale('log') ax.set_xlabel('Computational Complexity (GFLOPs) [Log Scale]', fontsize=10, fontweight='bold', labelpad=7, color='#2E2E2E') ax.set_ylabel('Gaze Estimation Error (MAE in Degrees) [Lower is Better]', fontsize=10, fontweight='bold', labelpad=7, color='#2E2E2E') ax.set_xlim(0.005, 600.0) ax.set_ylim(8.2, 3.5) ax.grid(True, which='both', ls=':', color='#DDDDDD', zorder=1) ax.tick_params(axis='both', colors='#555555', labelsize=8.5) for spine in ax.spines.values(): spine.set_edgecolor('#CCCCCC') spine.set_linewidth(0.7) # ── 7. LEGENDS ──────────────────────────────────────────────────────────────── # Left legend: Model classification group_handles = [] for g in ['ours', 'recent', 'teacher', 'prior']: st = GROUP_STYLE[g] mk = 'D' if g == 'ours' else 'o' group_handles.append(mlines.Line2D( [], [], color='none', marker=mk, markersize=7, markerfacecolor=st['color'], markeredgecolor=st['edgecolor'], markeredgewidth=1.1, label=st['label'], )) group_handles.append(mlines.Line2D( [], [], color='#888888', linestyle='--', linewidth=1.3, label='Current Pareto Frontier', )) leg1 = ax.legend( handles=group_handles, loc='lower left', title='Model Classification', fontsize=8, title_fontsize=8.5, frameon=True, facecolor='white', edgecolor='#CCCCCC', framealpha=0.93, borderpad=0.7, labelspacing=0.5, bbox_to_anchor=(0.01, 0.01), ) ax.add_artist(leg1) # Right legend: Bubble size sizes_demo = [0.18, 1.94, 11.72, 190.0] size_labels = ['0.18M (Ours)', '1.94M', '11.72M', '190.0M'] size_handles = [ plt.scatter([], [], s=bubble_area(sz) * 0.50, color='#C8C5BC', alpha=0.7, edgecolors='#5F5E5A', marker='o') for sz in sizes_demo ] leg2 = ax.legend( handles=size_handles, labels=size_labels, loc='lower right', title='Bubble Size (# Params)', fontsize=8, title_fontsize=8.5, frameon=True, facecolor='white', edgecolor='#CCCCCC', framealpha=0.93, labelspacing=1.2, borderpad=0.9, handletextpad=1.1, bbox_to_anchor=(0.99, 0.01), ) # ── 8. TITLE ────────────────────────────────────────────────────────────────── ax.set_title( 'Accuracy vs. Efficiency Trade-off: Gaze Estimation Benchmark', fontsize=11, fontweight='bold', color='#1A1A1A', pad=10, ) plt.tight_layout(pad=1.4) plt.savefig('/mnt/user-data/outputs/Figure_1_v2.png', dpi=300, bbox_inches='tight', facecolor=fig.get_facecolor()) print("Saved Figure_1_v2.png") plt.show()