diff --git a/taco_analysis/analyze_taco.py b/taco_analysis/analyze_taco.py
new file mode 100644
index 0000000000000000000000000000000000000000..ceb91461af5bf0ad4c61fadf9202f11cc48ae5ea
--- /dev/null
+++ b/taco_analysis/analyze_taco.py
@@ -0,0 +1,295 @@
+#!/usr/bin/env python
+"""Analyze the TACO dataset and generate summary plots.
+
+Usage:
+    python analyze_taco.py [--csv /path/to/taco_info.csv] [--output-dir /path/to/plots]
+"""
+
+import argparse
+from pathlib import Path
+
+import matplotlib
+matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+import numpy as np
+import pandas as pd
+import seaborn as sns
+
+sns.set_theme(style="whitegrid", font_scale=1.1)
+FIGSIZE = (12, 6)
+DPI = 150
+
+
+def plot_action_distribution(df: pd.DataFrame, out_dir: Path):
+    """Bar chart of sequences per action verb."""
+    fig, ax = plt.subplots(figsize=FIGSIZE)
+    counts = df["action"].value_counts().sort_values(ascending=True)
+    counts.plot.barh(ax=ax, color=sns.color_palette("viridis", len(counts)))
+    ax.set_xlabel("Number of sequences")
+    ax.set_title("Sequences per Action")
+    for i, v in enumerate(counts):
+        ax.text(v + 1, i, str(v), va="center", fontsize=9)
+    plt.tight_layout()
+    fig.savefig(out_dir / "action_distribution.png", dpi=DPI)
+    plt.close(fig)
+
+
+def plot_tool_distribution(df: pd.DataFrame, out_dir: Path):
+    """Bar chart of sequences per tool."""
+    fig, ax = plt.subplots(figsize=FIGSIZE)
+    counts = df["tool"].value_counts().sort_values(ascending=True)
+    counts.plot.barh(ax=ax, color=sns.color_palette("mako", len(counts)))
+    ax.set_xlabel("Number of sequences")
+    ax.set_title("Sequences per Tool")
+    for i, v in enumerate(counts):
+        ax.text(v + 1, i, str(v), va="center", fontsize=9)
+    plt.tight_layout()
+    fig.savefig(out_dir / "tool_distribution.png", dpi=DPI)
+    plt.close(fig)
+
+
+def plot_object_distribution(df: pd.DataFrame, out_dir: Path):
+    """Bar chart of sequences per object."""
+    fig, ax = plt.subplots(figsize=FIGSIZE)
+    counts = df["object"].value_counts().sort_values(ascending=True)
+    counts.plot.barh(ax=ax, color=sns.color_palette("rocket", len(counts)))
+    ax.set_xlabel("Number of sequences")
+    ax.set_title("Sequences per Object")
+    for i, v in enumerate(counts):
+        ax.text(v + 1, i, str(v), va="center", fontsize=9)
+    plt.tight_layout()
+    fig.savefig(out_dir / "object_distribution.png", dpi=DPI)
+    plt.close(fig)
+
+
+def plot_triplet_heatmap(df: pd.DataFrame, out_dir: Path):
+    """Heatmap of (action, object) combinations."""
+    pivot = df.groupby(["action", "object"]).size().reset_index(name="count")
+    pivot = pivot.pivot(index="action", columns="object", values="count").fillna(0).astype(int)
+    fig, ax = plt.subplots(figsize=(14, 8))
+    sns.heatmap(pivot, annot=True, fmt="d", cmap="YlOrRd", ax=ax, linewidths=0.5)
+    ax.set_title("Sequences per (Action, Object) Combination")
+    plt.tight_layout()
+    fig.savefig(out_dir / "action_object_heatmap.png", dpi=DPI)
+    plt.close(fig)
+
+
+def plot_frame_count_distribution(df: pd.DataFrame, out_dir: Path):
+    """Histogram of frame counts across sequences."""
+    valid = df[df["n_frames"] > 0]
+    if valid.empty:
+        return
+    fig, ax = plt.subplots(figsize=FIGSIZE)
+    ax.hist(valid["n_frames"], bins=50, color=sns.color_palette("deep")[0], edgecolor="white", alpha=0.8)
+    ax.axvline(valid["n_frames"].median(), color="red", linestyle="--", label=f'Median: {valid["n_frames"].median():.0f}')
+    ax.axvline(valid["n_frames"].mean(), color="orange", linestyle="--", label=f'Mean: {valid["n_frames"].mean():.0f}')
+    ax.set_xlabel("Number of frames")
+    ax.set_ylabel("Number of sequences")
+    ax.set_title("Frame Count Distribution")
+    ax.legend()
+    plt.tight_layout()
+    fig.savefig(out_dir / "frame_count_distribution.png", dpi=DPI)
+    plt.close(fig)
+
+
+def plot_duration_distribution(df: pd.DataFrame, out_dir: Path):
+    """Histogram of sequence durations."""
+    valid = df[df["duration_s"] > 0]
+    if valid.empty:
+        return
+    fig, ax = plt.subplots(figsize=FIGSIZE)
+    ax.hist(valid["duration_s"], bins=50, color=sns.color_palette("deep")[1], edgecolor="white", alpha=0.8)
+    ax.axvline(valid["duration_s"].median(), color="red", linestyle="--", label=f'Median: {valid["duration_s"].median():.1f}s')
+    ax.axvline(valid["duration_s"].mean(), color="orange", linestyle="--", label=f'Mean: {valid["duration_s"].mean():.1f}s')
+    ax.set_xlabel("Duration (seconds)")
+    ax.set_ylabel("Number of sequences")
+    ax.set_title("Sequence Duration Distribution")
+    ax.legend()
+    plt.tight_layout()
+    fig.savefig(out_dir / "duration_distribution.png", dpi=DPI)
+    plt.close(fig)
+
+
+def plot_duration_by_action(df: pd.DataFrame, out_dir: Path):
+    """Box plot of duration per action."""
+    valid = df[df["duration_s"] > 0]
+    if valid.empty:
+        return
+    order = valid.groupby("action")["duration_s"].median().sort_values(ascending=False).index
+    fig, ax = plt.subplots(figsize=FIGSIZE)
+    sns.boxplot(data=valid, x="action", y="duration_s", order=order, ax=ax, palette="viridis")
+    ax.set_xlabel("Action")
+    ax.set_ylabel("Duration (seconds)")
+    ax.set_title("Sequence Duration by Action")
+    plt.xticks(rotation=45, ha="right")
+    plt.tight_layout()
+    fig.savefig(out_dir / "duration_by_action.png", dpi=DPI)
+    plt.close(fig)
+
+
+def plot_modality_completeness(df: pd.DataFrame, out_dir: Path):
+    """Bar chart of modality availability."""
+    modality_cols = [
+        "has_egocentric_rgb", "has_egocentric_depth", "has_hand_poses",
+        "has_object_poses", "has_segmentation", "has_alloc_cam_params",
+        "has_ego_cam_params",
+    ]
+    labels = [c.replace("has_", "").replace("_", " ").title() for c in modality_cols]
+    counts = [df[c].sum() for c in modality_cols]
+    pcts = [100 * c / len(df) for c in counts]
+
+    fig, ax = plt.subplots(figsize=FIGSIZE)
+    bars = ax.barh(labels, pcts, color=sns.color_palette("Greens_d", len(labels)))
+    ax.set_xlabel("Availability (%)")
+    ax.set_title("Modality Availability Across Sequences")
+    ax.set_xlim(0, 105)
+    for i, (p, c) in enumerate(zip(pcts, counts)):
+        ax.text(p + 0.5, i, f"{p:.1f}% ({c}/{len(df)})", va="center", fontsize=9)
+    plt.tight_layout()
+    fig.savefig(out_dir / "modality_completeness.png", dpi=DPI)
+    plt.close(fig)
+
+
+def plot_sequences_per_triplet(df: pd.DataFrame, out_dir: Path):
+    """Histogram of how many sequences each triplet has."""
+    triplet_counts = df["triplet"].value_counts()
+    fig, ax = plt.subplots(figsize=FIGSIZE)
+    ax.hist(triplet_counts, bins=range(1, triplet_counts.max() + 2), color=sns.color_palette("deep")[2],
+            edgecolor="white", alpha=0.8, align="left")
+    ax.set_xlabel("Sequences per triplet")
+    ax.set_ylabel("Number of triplets")
+    ax.set_title(f"Sequences per Triplet ({df['triplet'].nunique()} unique triplets)")
+    ax.axvline(triplet_counts.median(), color="red", linestyle="--",
+               label=f"Median: {triplet_counts.median():.0f}")
+    ax.legend()
+    plt.tight_layout()
+    fig.savefig(out_dir / "sequences_per_triplet.png", dpi=DPI)
+    plt.close(fig)
+
+
+def plot_date_timeline(df: pd.DataFrame, out_dir: Path):
+    """Sequences collected over time."""
+    dates = pd.to_datetime(df["date"], format="%Y%m%d", errors="coerce")
+    valid = dates.dropna()
+    if valid.empty:
+        return
+    fig, ax = plt.subplots(figsize=FIGSIZE)
+    daily = valid.groupby(valid.dt.date).size()
+    daily.plot(kind="bar", ax=ax, color=sns.color_palette("deep")[3], width=0.8)
+    ax.set_xlabel("Date")
+    ax.set_ylabel("Sequences recorded")
+    ax.set_title("Data Collection Timeline")
+    # Reduce x-tick clutter
+    n_ticks = len(daily)
+    if n_ticks > 20:
+        step = max(1, n_ticks // 15)
+        ticks = ax.get_xticks()
+        ax.set_xticks(ticks[::step])
+    plt.xticks(rotation=45, ha="right")
+    plt.tight_layout()
+    fig.savefig(out_dir / "collection_timeline.png", dpi=DPI)
+    plt.close(fig)
+
+
+def plot_frames_by_action(df: pd.DataFrame, out_dir: Path):
+    """Box plot of frame counts per action."""
+    valid = df[df["n_frames"] > 0]
+    if valid.empty:
+        return
+    order = valid.groupby("action")["n_frames"].median().sort_values(ascending=False).index
+    fig, ax = plt.subplots(figsize=FIGSIZE)
+    sns.boxplot(data=valid, x="action", y="n_frames", order=order, ax=ax, palette="mako")
+    ax.set_xlabel("Action")
+    ax.set_ylabel("Frame count")
+    ax.set_title("Frame Count by Action")
+    plt.xticks(rotation=45, ha="right")
+    plt.tight_layout()
+    fig.savefig(out_dir / "frames_by_action.png", dpi=DPI)
+    plt.close(fig)
+
+
+def print_summary(df: pd.DataFrame):
+    """Print text summary to console."""
+    print(f"\n{'='*70}")
+    print("TACO DATASET ANALYSIS SUMMARY")
+    print(f"{'='*70}")
+    print(f"Total sequences:     {len(df)}")
+    print(f"Unique triplets:     {df['triplet'].nunique()}")
+    print(f"Unique actions:      {df['action'].nunique()}: {sorted(df['action'].unique())}")
+    print(f"Unique tools:        {df['tool'].nunique()}: {sorted(df['tool'].unique())}")
+    print(f"Unique objects:      {df['object'].nunique()}: {sorted(df['object'].unique())}")
+    print(f"All 12 cameras:      {(df['n_allocentric_cameras'] == 12).sum()}/{len(df)}")
+    complete = df["all_modalities_complete"].sum()
+    print(f"Fully complete:      {complete}/{len(df)} ({100*complete/len(df):.1f}%)")
+
+    valid = df[df["n_frames"] > 0]
+    if len(valid) > 0:
+        print(f"\nVideo statistics ({len(valid)} sequences with video info):")
+        print(f"  FPS:      {valid['fps'].mode().iloc[0]}")
+        print(f"  Frames:   min={valid['n_frames'].min()}, median={valid['n_frames'].median():.0f}, "
+              f"mean={valid['n_frames'].mean():.0f}, max={valid['n_frames'].max()}")
+        print(f"  Duration: min={valid['duration_s'].min():.1f}s, median={valid['duration_s'].median():.1f}s, "
+              f"mean={valid['duration_s'].mean():.1f}s, max={valid['duration_s'].max():.1f}s")
+        print(f"  Total:    {valid['duration_s'].sum()/3600:.1f} hours")
+
+    # Top triplets
+    top = df["triplet"].value_counts().head(10)
+    print(f"\nTop 10 triplets by sequence count:")
+    for triplet, count in top.items():
+        print(f"  {triplet}: {count}")
+
+    print(f"{'='*70}")
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Analyze TACO dataset and generate plots")
+    parser.add_argument(
+        "--csv", type=Path,
+        default=Path(__file__).resolve().parent.parent / "taco_dataset" / "taco_info.csv",
+        help="Path to taco_info.csv",
+    )
+    parser.add_argument(
+        "--output-dir", type=Path,
+        default=Path(__file__).resolve().parent / "plots",
+        help="Directory to save plots",
+    )
+    args = parser.parse_args()
+
+    df = pd.read_csv(args.csv)
+    out_dir = args.output_dir
+    out_dir.mkdir(parents=True, exist_ok=True)
+
+    print(f"Loaded {len(df)} sequences from {args.csv}")
+    print(f"Saving plots to {out_dir}/")
+
+    print_summary(df)
+
+    print("\nGenerating plots...")
+    plot_action_distribution(df, out_dir)
+    print("  action_distribution.png")
+    plot_tool_distribution(df, out_dir)
+    print("  tool_distribution.png")
+    plot_object_distribution(df, out_dir)
+    print("  object_distribution.png")
+    plot_triplet_heatmap(df, out_dir)
+    print("  action_object_heatmap.png")
+    plot_frame_count_distribution(df, out_dir)
+    print("  frame_count_distribution.png")
+    plot_duration_distribution(df, out_dir)
+    print("  duration_distribution.png")
+    plot_duration_by_action(df, out_dir)
+    print("  duration_by_action.png")
+    plot_modality_completeness(df, out_dir)
+    print("  modality_completeness.png")
+    plot_sequences_per_triplet(df, out_dir)
+    print("  sequences_per_triplet.png")
+    plot_date_timeline(df, out_dir)
+    print("  collection_timeline.png")
+    plot_frames_by_action(df, out_dir)
+    print("  frames_by_action.png")
+
+    print(f"\nDone! {11} plots saved to {out_dir}/")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/taco_analysis/plots/action_distribution.png b/taco_analysis/plots/action_distribution.png
new file mode 100644
index 0000000000000000000000000000000000000000..5460a6028862c0f5720ddbf22dcf885107737152
--- /dev/null
+++ b/taco_analysis/plots/action_distribution.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:97fefdc6a855a740fcc5804f37c20d8817b5b8142dc3d782d0d8b894b258f54d
+size 72066
diff --git a/taco_analysis/plots/action_object_heatmap.png b/taco_analysis/plots/action_object_heatmap.png
new file mode 100644
index 0000000000000000000000000000000000000000..43d83b8a56578c65ac6b56fa29a108798914d4b1
--- /dev/null
+++ b/taco_analysis/plots/action_object_heatmap.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:602d69c4d26ebe309508d2f3020a8bcc3691d68cd23cd57134933110d34e521b
+size 140983
diff --git a/taco_analysis/plots/collection_timeline.png b/taco_analysis/plots/collection_timeline.png
new file mode 100644
index 0000000000000000000000000000000000000000..1972e62c66d831e03767a02109aa6d43cbf0b772
--- /dev/null
+++ b/taco_analysis/plots/collection_timeline.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:418221a4e461c2ee71463596cd05ad52383f399aa7331d0ff449802d1ac006ad
+size 60167
diff --git a/taco_analysis/plots/duration_by_action.png b/taco_analysis/plots/duration_by_action.png
new file mode 100644
index 0000000000000000000000000000000000000000..67640ba5de5793cd22281a62ff3768ab8cd75622
--- /dev/null
+++ b/taco_analysis/plots/duration_by_action.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:675b70ddc10dd00dc6a204b6cd0af98ab8695b944fb53991abeea398ce861e3a
+size 107244
diff --git a/taco_analysis/plots/duration_distribution.png b/taco_analysis/plots/duration_distribution.png
new file mode 100644
index 0000000000000000000000000000000000000000..5ea0a8b1c9ab378eba79fb5279a4486b2117da69
--- /dev/null
+++ b/taco_analysis/plots/duration_distribution.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:941b02c458d831748d0619d3c796c1755d72f22d7aaf894ad56662142d080621
+size 55481
diff --git a/taco_analysis/plots/epipolar/brushbrushcup_20231026_002_f0000_src4.png b/taco_analysis/plots/epipolar/brushbrushcup_20231026_002_f0000_src4.png
new file mode 100644
index 0000000000000000000000000000000000000000..9280574c2fed68d43164d076e9627c6f95215bf7
--- /dev/null
+++ b/taco_analysis/plots/epipolar/brushbrushcup_20231026_002_f0000_src4.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:27e4b2e6647667aad7a03675f55cea53f36bca2c85a003238dd8501c0cec7ab7
+size 2861117
diff --git a/taco_analysis/plots/epipolar/brushbrushcup_20231026_006_f0000_src4.png b/taco_analysis/plots/epipolar/brushbrushcup_20231026_006_f0000_src4.png
new file mode 100644
index 0000000000000000000000000000000000000000..48184e711f67ef9bf5e9ea260b04439796a58f9a
--- /dev/null
+++ b/taco_analysis/plots/epipolar/brushbrushcup_20231026_006_f0000_src4.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e082d500ac2ef72635c684569d15974927086d2d56ed0348fbf5bd5d6fa4ca6b
+size 2910576
diff --git a/taco_analysis/plots/epipolar/brushbrushcup_20231026_006_f0183_src4.png b/taco_analysis/plots/epipolar/brushbrushcup_20231026_006_f0183_src4.png
new file mode 100644
index 0000000000000000000000000000000000000000..62f58cb4e0050b3ec2e542cc014faaf717b8af16
--- /dev/null
+++ b/taco_analysis/plots/epipolar/brushbrushcup_20231026_006_f0183_src4.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b019d1a78cf37810a83e5a253d7093b0b79918c7265eeb346949c457f6f974c7
+size 3185512
diff --git a/taco_analysis/plots/epipolar/brushbrushhelmet_20231027_007_f0000_src5.png b/taco_analysis/plots/epipolar/brushbrushhelmet_20231027_007_f0000_src5.png
new file mode 100644
index 0000000000000000000000000000000000000000..7fb8452950cb3054a87e6e9644c95f68d526ffab
--- /dev/null
+++ b/taco_analysis/plots/epipolar/brushbrushhelmet_20231027_007_f0000_src5.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2972ecb49d8810aa1ef1eb020baf606d9e7218646683d1a12a046fca03b1c13e
+size 2956336
diff --git a/taco_analysis/plots/epipolar/brushbrushkettle_20231006_169_f0000_src1.png b/taco_analysis/plots/epipolar/brushbrushkettle_20231006_169_f0000_src1.png
new file mode 100644
index 0000000000000000000000000000000000000000..5f48ec2292162007b4554973aae42900a963cc2a
--- /dev/null
+++ b/taco_analysis/plots/epipolar/brushbrushkettle_20231006_169_f0000_src1.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:54a0869dbc3562e87e9d12d1544133c229a9d94d4833056f6e80650ba99f294c
+size 2989448
diff --git a/taco_analysis/plots/epipolar/brushbrushkettle_20231006_169_f0000_src2.png b/taco_analysis/plots/epipolar/brushbrushkettle_20231006_169_f0000_src2.png
new file mode 100644
index 0000000000000000000000000000000000000000..c3f95d6f7b170d5f5e956050940e6e930b00af9c
--- /dev/null
+++ b/taco_analysis/plots/epipolar/brushbrushkettle_20231006_169_f0000_src2.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9e8aa717a593ac0b70b66f03bbc94acaab9b683fd383a9b3717e46d59cccacf2
+size 3092834
diff --git a/taco_analysis/plots/epipolar/brushbrushkettle_20231006_169_f0137_src1.png b/taco_analysis/plots/epipolar/brushbrushkettle_20231006_169_f0137_src1.png
new file mode 100644
index 0000000000000000000000000000000000000000..505cf41176fe67cedd200af48ee8dcc341a9d506
--- /dev/null
+++ b/taco_analysis/plots/epipolar/brushbrushkettle_20231006_169_f0137_src1.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8c9cbedf3cc74b7463135cb032d1e6d0b079dc2cc61846bea3034bc578a0a91a
+size 3320620
diff --git a/taco_analysis/plots/epipolar/brushbrushkettle_20231103_161_f0000_src3.png b/taco_analysis/plots/epipolar/brushbrushkettle_20231103_161_f0000_src3.png
new file mode 100644
index 0000000000000000000000000000000000000000..64e67f03f685cdedc1d17a1371264a9da7f6c26b
--- /dev/null
+++ b/taco_analysis/plots/epipolar/brushbrushkettle_20231103_161_f0000_src3.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1b07d2b86dd2a87ad3f86e5ec53ec6a61fb5e864958f6da442a7d60528735253
+size 2968570
diff --git a/taco_analysis/plots/epipolar/brushbrushkettle_20231103_161_f0134_src3.png b/taco_analysis/plots/epipolar/brushbrushkettle_20231103_161_f0134_src3.png
new file mode 100644
index 0000000000000000000000000000000000000000..9f6017abdfda62ca12637c163f8cfcc57c11aa47
--- /dev/null
+++ b/taco_analysis/plots/epipolar/brushbrushkettle_20231103_161_f0134_src3.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:731281700fe1807daa68cba9517c003c0adfec83c25848698f923a69622644f5
+size 3260430
diff --git a/taco_analysis/plots/epipolar/brushbrushplate_20230928_044_f0000_src6.png b/taco_analysis/plots/epipolar/brushbrushplate_20230928_044_f0000_src6.png
new file mode 100644
index 0000000000000000000000000000000000000000..3b4b68cbf406971703105f4e58efca5420c85722
--- /dev/null
+++ b/taco_analysis/plots/epipolar/brushbrushplate_20230928_044_f0000_src6.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ee17202062b8ac5eb8c9499185b60c15e0b639eb424e15e6ed2bce195f176316
+size 3013159
diff --git a/taco_analysis/plots/epipolar/brushbrushplate_20230929_039_f0000_src8.png b/taco_analysis/plots/epipolar/brushbrushplate_20230929_039_f0000_src8.png
new file mode 100644
index 0000000000000000000000000000000000000000..fbae8784dedba6ebb46feb30aa8d41bac835eeb8
--- /dev/null
+++ b/taco_analysis/plots/epipolar/brushbrushplate_20230929_039_f0000_src8.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:635c993feb645c74177f41d42cb1e06c12422cab445b2a2ccd3e39232dc5a071
+size 3140295
diff --git a/taco_analysis/plots/epipolar/brushbrushplate_20231005_184_f0000_src2.png b/taco_analysis/plots/epipolar/brushbrushplate_20231005_184_f0000_src2.png
new file mode 100644
index 0000000000000000000000000000000000000000..2d120b489b0d4669c66ea33e9a681a0251002360
--- /dev/null
+++ b/taco_analysis/plots/epipolar/brushbrushplate_20231005_184_f0000_src2.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:25c34bcce11a52f2b3a2fcf35dbbb434e9393c7890ce50bfb921f913caad0820
+size 3111232
diff --git a/taco_analysis/plots/epipolar/brushbrushplate_20231005_184_f0076_src2.png b/taco_analysis/plots/epipolar/brushbrushplate_20231005_184_f0076_src2.png
new file mode 100644
index 0000000000000000000000000000000000000000..bb26a504edd18722e5c8189d0a7129ab110c2a52
--- /dev/null
+++ b/taco_analysis/plots/epipolar/brushbrushplate_20231005_184_f0076_src2.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ba3c695109ea2e864d13da85b06cb72efd524f930f166ed433051b8092128222
+size 3412136
diff --git a/taco_analysis/plots/epipolar/brushbrushteapot_20231006_227_f0000_src3.png b/taco_analysis/plots/epipolar/brushbrushteapot_20231006_227_f0000_src3.png
new file mode 100644
index 0000000000000000000000000000000000000000..a0dfd693919f1b90258dd1de1683087ac7a63e41
--- /dev/null
+++ b/taco_analysis/plots/epipolar/brushbrushteapot_20231006_227_f0000_src3.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b1b40b411d2f6900c22c3b0b016413cce46690d1b24d10533e863dfc7ff6aa12
+size 3089194
diff --git a/taco_analysis/plots/epipolar/brusheraserbowl_20231103_027_f0000_src0.png b/taco_analysis/plots/epipolar/brusheraserbowl_20231103_027_f0000_src0.png
new file mode 100644
index 0000000000000000000000000000000000000000..879eee895b3fbc9dfcf4739ee4712e75d640e8d2
--- /dev/null
+++ b/taco_analysis/plots/epipolar/brusheraserbowl_20231103_027_f0000_src0.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:24405e6e577233871b5919b688dd42652849cdcbc7f7e641ef53fc09c8873060
+size 3110612
diff --git a/taco_analysis/plots/epipolar/brusheraserbowl_20231103_027_f0099_src0.png b/taco_analysis/plots/epipolar/brusheraserbowl_20231103_027_f0099_src0.png
new file mode 100644
index 0000000000000000000000000000000000000000..047299f93a89461526cf97c7d748df7a751d8755
--- /dev/null
+++ b/taco_analysis/plots/epipolar/brusheraserbowl_20231103_027_f0099_src0.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ec70df95fd1d5077a657eccfa7f9813514a6bfad97698d252f1c117137475d04
+size 3327554
diff --git a/taco_analysis/plots/epipolar/brusheraserbowl_20231103_031_f0000_src7.png b/taco_analysis/plots/epipolar/brusheraserbowl_20231103_031_f0000_src7.png
new file mode 100644
index 0000000000000000000000000000000000000000..06a00ecf40200ce8369856d663a933da1d5812d4
--- /dev/null
+++ b/taco_analysis/plots/epipolar/brusheraserbowl_20231103_031_f0000_src7.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6cb2638e0596edba282ff4360acb3178b30f3bae02a1df814fa2cb8d798739ad
+size 3168583
diff --git a/taco_analysis/plots/epipolar/brusheraserbowl_20231103_045_f0000_src1.png b/taco_analysis/plots/epipolar/brusheraserbowl_20231103_045_f0000_src1.png
new file mode 100644
index 0000000000000000000000000000000000000000..49996720cc9ae04704a418793600d7879b795896
--- /dev/null
+++ b/taco_analysis/plots/epipolar/brusheraserbowl_20231103_045_f0000_src1.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4e1b4261bfeff1ac2ba5d5b35383a59ac0cd294999c1d9fb8ba8c1273bd830ea
+size 3148248
diff --git a/taco_analysis/plots/epipolar/brusheraserkettle_20231103_110_f0000_src2.png b/taco_analysis/plots/epipolar/brusheraserkettle_20231103_110_f0000_src2.png
new file mode 100644
index 0000000000000000000000000000000000000000..eab78602353bbda93f9937b2524fb4782c6b1c7f
--- /dev/null
+++ b/taco_analysis/plots/epipolar/brusheraserkettle_20231103_110_f0000_src2.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2f879875212ad252707eb174caff08d3f81b958ce232994f1a934602802a45d1
+size 3161093
diff --git a/taco_analysis/plots/epipolar/brusheraserkettle_20231103_110_f0103_src2.png b/taco_analysis/plots/epipolar/brusheraserkettle_20231103_110_f0103_src2.png
new file mode 100644
index 0000000000000000000000000000000000000000..0407d9b0ff552729a4bc41e870d92c5a8cfd601d
--- /dev/null
+++ b/taco_analysis/plots/epipolar/brusheraserkettle_20231103_110_f0103_src2.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:520467532b586358aa58251be5eac1eb5518cb001901ddd6c6d6d07f55736a19
+size 3470726
diff --git a/taco_analysis/plots/epipolar/brusheraserpan_20231103_001_f0000_src0.png b/taco_analysis/plots/epipolar/brusheraserpan_20231103_001_f0000_src0.png
new file mode 100644
index 0000000000000000000000000000000000000000..cc4df42c8593cead5bb0f43ef00e7b1add3378ab
--- /dev/null
+++ b/taco_analysis/plots/epipolar/brusheraserpan_20231103_001_f0000_src0.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:19f036aad23008c4d439f1e1fe304bce86cc6a0f4800fef9da3be85d93e56d0d
+size 3120302
diff --git a/taco_analysis/plots/epipolar/brusheraserplate_20231027_049_f0000_src9.png b/taco_analysis/plots/epipolar/brusheraserplate_20231027_049_f0000_src9.png
new file mode 100644
index 0000000000000000000000000000000000000000..f328a30ce619ff556662ce566c8012dd22abea93
--- /dev/null
+++ b/taco_analysis/plots/epipolar/brusheraserplate_20231027_049_f0000_src9.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0d123716c206bdc032bd7930f40658673a6c3d8649b8584fba6e40cd3ad7203b
+size 3030657
diff --git a/taco_analysis/plots/epipolar/brushrollerplate_20231019_241_f0000_src5.png b/taco_analysis/plots/epipolar/brushrollerplate_20231019_241_f0000_src5.png
new file mode 100644
index 0000000000000000000000000000000000000000..ae75f793634493e06420ab78a2dc457c34ff605c
--- /dev/null
+++ b/taco_analysis/plots/epipolar/brushrollerplate_20231019_241_f0000_src5.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b824bd983eab262c3004bb0ad19fb5c48ef353429aeb58f73dd6cc7742228c2e
+size 2977552
diff --git a/taco_analysis/plots/epipolar/cutspatulabowl_20231027_067_f0000_src6.png b/taco_analysis/plots/epipolar/cutspatulabowl_20231027_067_f0000_src6.png
new file mode 100644
index 0000000000000000000000000000000000000000..0d2dc1fafdade246990a6fec6f92028afa5326c8
--- /dev/null
+++ b/taco_analysis/plots/epipolar/cutspatulabowl_20231027_067_f0000_src6.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fad65fc69637dcd94fe0166c6d9169403fffe5577560af3d208d84897b8e5f20
+size 2979308
diff --git a/taco_analysis/plots/epipolar/dustbrushkettle_20231103_160_f0000_src3.png b/taco_analysis/plots/epipolar/dustbrushkettle_20231103_160_f0000_src3.png
new file mode 100644
index 0000000000000000000000000000000000000000..ea3ed88cdc38c8f32b662efdc03093bd2e875098
--- /dev/null
+++ b/taco_analysis/plots/epipolar/dustbrushkettle_20231103_160_f0000_src3.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:75e87d9a2c80b1a90aa89693c95bc321857e88e41432c47504c4566af416268d
+size 2980652
diff --git a/taco_analysis/plots/epipolar/measurerulerbowl_20231104_088_f0000_src4.png b/taco_analysis/plots/epipolar/measurerulerbowl_20231104_088_f0000_src4.png
new file mode 100644
index 0000000000000000000000000000000000000000..7f2921f623b079c68b705d93f0b9a1d218f79f9b
--- /dev/null
+++ b/taco_analysis/plots/epipolar/measurerulerbowl_20231104_088_f0000_src4.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8f569010d7181e4b63624be5b9e30c4eed885a071892d8d7ac0f3b28310a739e
+size 2841991
diff --git a/taco_analysis/plots/epipolar/measurerulerbowl_20231104_088_f0135_src4.png b/taco_analysis/plots/epipolar/measurerulerbowl_20231104_088_f0135_src4.png
new file mode 100644
index 0000000000000000000000000000000000000000..286ac354bb7b81f24edb047f20d81b0264ea6c63
--- /dev/null
+++ b/taco_analysis/plots/epipolar/measurerulerbowl_20231104_088_f0135_src4.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8a2b4e5aaf6c9c2fe7e6c2bcfa79e6c43adaa413027d4bf1eb58443d0afca760
+size 3117422
diff --git a/taco_analysis/plots/epipolar/measurerulerplate_20231104_075_f0000_src1.png b/taco_analysis/plots/epipolar/measurerulerplate_20231104_075_f0000_src1.png
new file mode 100644
index 0000000000000000000000000000000000000000..290ddfde8b9c94609b07967a34d83a8370a2ec88
--- /dev/null
+++ b/taco_analysis/plots/epipolar/measurerulerplate_20231104_075_f0000_src1.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c0365540aa6a8af9051143ce059089b3b1d3fd9053dae04eaece5a7d8184463e
+size 3071365
diff --git a/taco_analysis/plots/epipolar/putinspatulapan_20231020_038_f0000_src9.png b/taco_analysis/plots/epipolar/putinspatulapan_20231020_038_f0000_src9.png
new file mode 100644
index 0000000000000000000000000000000000000000..cbaaf134c9822dbc5ee6e1ef734c3abea1d52fba
--- /dev/null
+++ b/taco_analysis/plots/epipolar/putinspatulapan_20231020_038_f0000_src9.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4ecffce415a78750938bc0ecd8e2b009a8aebe2ea61c954cc0552b7951ad0c26
+size 2836866
diff --git a/taco_analysis/plots/epipolar/putoutbowlpan_20231015_091_f0000_src0.png b/taco_analysis/plots/epipolar/putoutbowlpan_20231015_091_f0000_src0.png
new file mode 100644
index 0000000000000000000000000000000000000000..26d07eb94e41a0fc0a0dd822dc3d15f68c2d8a20
--- /dev/null
+++ b/taco_analysis/plots/epipolar/putoutbowlpan_20231015_091_f0000_src0.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ee0a4da158b100c1a562e5b548e22a972595c3b28409130512d24310aa4f21d9
+size 2937495
diff --git a/taco_analysis/plots/epipolar/scrapeoffknifeplate_20231015_210_f0000_src0.png b/taco_analysis/plots/epipolar/scrapeoffknifeplate_20231015_210_f0000_src0.png
new file mode 100644
index 0000000000000000000000000000000000000000..b4dfbae7dd32ee92404427b2c7123bb17c63a6bb
--- /dev/null
+++ b/taco_analysis/plots/epipolar/scrapeoffknifeplate_20231015_210_f0000_src0.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:08348e31bf9bad0cc2cb350beb6e6a376274f41a85d71bb2018f84af50220388
+size 2827718
diff --git a/taco_analysis/plots/epipolar/scrapeoffknifeplate_20231015_210_f0150_src0.png b/taco_analysis/plots/epipolar/scrapeoffknifeplate_20231015_210_f0150_src0.png
new file mode 100644
index 0000000000000000000000000000000000000000..364d8bf1628d5a0fe58e79e7a0823666e7ffd7ea
--- /dev/null
+++ b/taco_analysis/plots/epipolar/scrapeoffknifeplate_20231015_210_f0150_src0.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a53d4d96df08bff44387b0a635b094f017968fa481a68ed3b9a8d4acb639d36a
+size 3159139
diff --git a/taco_analysis/plots/epipolar/screwscrewdriverbox_20231102_047_f0000_src2.png b/taco_analysis/plots/epipolar/screwscrewdriverbox_20231102_047_f0000_src2.png
new file mode 100644
index 0000000000000000000000000000000000000000..89a76b61063acf6cc4f1aa6f4b25cd2b9b630b19
--- /dev/null
+++ b/taco_analysis/plots/epipolar/screwscrewdriverbox_20231102_047_f0000_src2.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3095119e2e213c68ca813a157d74cfc6bec1e0bebe98dcad017fb3b6d195d2b8
+size 3238337
diff --git a/taco_analysis/plots/epipolar/skimoffspoonplate_20231024_069_f0000_src1.png b/taco_analysis/plots/epipolar/skimoffspoonplate_20231024_069_f0000_src1.png
new file mode 100644
index 0000000000000000000000000000000000000000..08c06e77697a621997911c9c06286e5d3adb2b3a
--- /dev/null
+++ b/taco_analysis/plots/epipolar/skimoffspoonplate_20231024_069_f0000_src1.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:75fcef739ee4e53683a503ae2836f2fb90797417ddcae8163daad3602d94d52e
+size 2957994
diff --git a/taco_analysis/plots/epipolar/skimoffspoonplate_20231024_069_f0203_src1.png b/taco_analysis/plots/epipolar/skimoffspoonplate_20231024_069_f0203_src1.png
new file mode 100644
index 0000000000000000000000000000000000000000..6dd302ac0b979461a599930f0ba5b9b69cd2f226
--- /dev/null
+++ b/taco_analysis/plots/epipolar/skimoffspoonplate_20231024_069_f0203_src1.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ea6310b4b308d57173389732ebdc4dc3abbe90980dff36737527207148ea7053
+size 3243074
diff --git a/taco_analysis/plots/epipolar/smeargluegunbox_20231010_192_f0000_src4.png b/taco_analysis/plots/epipolar/smeargluegunbox_20231010_192_f0000_src4.png
new file mode 100644
index 0000000000000000000000000000000000000000..bdad1ba7987752ced7f36b16b9187dc5fe51b458
--- /dev/null
+++ b/taco_analysis/plots/epipolar/smeargluegunbox_20231010_192_f0000_src4.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1ffdb287c3cfc66e4ea3d7848fd2c4f700fdaf7c6f00495c606f4cef9fa2d5f1
+size 2814394
diff --git a/taco_analysis/plots/epipolar/stir-fryspatulaplate_20231104_127_f0000_src7.png b/taco_analysis/plots/epipolar/stir-fryspatulaplate_20231104_127_f0000_src7.png
new file mode 100644
index 0000000000000000000000000000000000000000..22911fd0d5231a9eacb0873426165612d8d3df56
--- /dev/null
+++ b/taco_analysis/plots/epipolar/stir-fryspatulaplate_20231104_127_f0000_src7.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2c2bf3688f39f77450750c6fe0149a559ad61ca0438981266416dc0d01d976b3
+size 2973804
diff --git a/taco_analysis/plots/epipolar/stirspatulabowl_20231024_111_f0000_src8.png b/taco_analysis/plots/epipolar/stirspatulabowl_20231024_111_f0000_src8.png
new file mode 100644
index 0000000000000000000000000000000000000000..8de5a98342e3fbfed169ab0bb061ea01a2dd727c
--- /dev/null
+++ b/taco_analysis/plots/epipolar/stirspatulabowl_20231024_111_f0000_src8.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:76c9926cabddecb35ec7988df688c8a461bd7c022d9dd29da6a789dc2ca0002a
+size 2961108
diff --git a/taco_analysis/plots/epipolar/stirspoonpan_20231020_079_f0000_src3.png b/taco_analysis/plots/epipolar/stirspoonpan_20231020_079_f0000_src3.png
new file mode 100644
index 0000000000000000000000000000000000000000..09a0f830764c66b5ba8a0ef5509c6790c62d9a7e
--- /dev/null
+++ b/taco_analysis/plots/epipolar/stirspoonpan_20231020_079_f0000_src3.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ffed82e26fe902060f0a512aa88a0239a8ac61c1fb1458cfb48aaa2303bcd6e4
+size 2902155
diff --git a/taco_analysis/plots/epipolar/stirspoonpan_20231020_079_f0168_src3.png b/taco_analysis/plots/epipolar/stirspoonpan_20231020_079_f0168_src3.png
new file mode 100644
index 0000000000000000000000000000000000000000..20b3c633e995f0f4b381402e1a611339377a066d
--- /dev/null
+++ b/taco_analysis/plots/epipolar/stirspoonpan_20231020_079_f0168_src3.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:382dba44b08ea922fbfe6a6c9cc694592eb756ed25c2a5b78c801d32f2ed4ee4
+size 3166123
diff --git a/taco_analysis/plots/frame_count_distribution.png b/taco_analysis/plots/frame_count_distribution.png
new file mode 100644
index 0000000000000000000000000000000000000000..988b2efbc885c098493371029297e0047629c3cf
--- /dev/null
+++ b/taco_analysis/plots/frame_count_distribution.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:33d55276b47b4f4dc669ce47a855a1f685ed287dd3f28bdf064a29f7b4889f76
+size 50557
diff --git a/taco_analysis/plots/frames_by_action.png b/taco_analysis/plots/frames_by_action.png
new file mode 100644
index 0000000000000000000000000000000000000000..b42c3dbd3004c6027addcbc0fb88296167461735
--- /dev/null
+++ b/taco_analysis/plots/frames_by_action.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:740f64772ba20ef279db5960b3d6eaaee2652006c53d306db577ded61aa993ac
+size 103786
diff --git a/taco_analysis/plots/mesh_overlay/dustrollerpan_20231019_214_f0000.png b/taco_analysis/plots/mesh_overlay/dustrollerpan_20231019_214_f0000.png
new file mode 100644
index 0000000000000000000000000000000000000000..bdefab76d2d93e38df6945395d42b655b4db6e77
--- /dev/null
+++ b/taco_analysis/plots/mesh_overlay/dustrollerpan_20231019_214_f0000.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:24fe16226087c83219124d656fa50529949ba5cf6107ad80fbd4f83d3030c7ce
+size 2820098
diff --git a/taco_analysis/plots/mesh_overlay/dustrollerpan_20231019_214_f0116.png b/taco_analysis/plots/mesh_overlay/dustrollerpan_20231019_214_f0116.png
new file mode 100644
index 0000000000000000000000000000000000000000..75ec123fdd1f8c5bd2752e53dca01b3853d083e4
--- /dev/null
+++ b/taco_analysis/plots/mesh_overlay/dustrollerpan_20231019_214_f0116.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8c1d99fe504e598a89dd061bfca2117835961720d9a4f5a207b04b96aa5a9240
+size 3308931
diff --git a/taco_analysis/plots/mesh_overlay/emptybowlkettle_20231019_008_f0000.png b/taco_analysis/plots/mesh_overlay/emptybowlkettle_20231019_008_f0000.png
new file mode 100644
index 0000000000000000000000000000000000000000..c8d0b7bc87334c19e3865016b6f7ec6cc7f500c6
--- /dev/null
+++ b/taco_analysis/plots/mesh_overlay/emptybowlkettle_20231019_008_f0000.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:92c695a4090988201a464145f3877e46d2da8953586ed4733c5064ebedaaf382
+size 2718398
diff --git a/taco_analysis/plots/mesh_overlay/emptybowlkettle_20231019_008_f0149.png b/taco_analysis/plots/mesh_overlay/emptybowlkettle_20231019_008_f0149.png
new file mode 100644
index 0000000000000000000000000000000000000000..430d0c2b644857553e956b3065f0643fa8ceb0f7
--- /dev/null
+++ b/taco_analysis/plots/mesh_overlay/emptybowlkettle_20231019_008_f0149.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8bc32f841b911a27d9f769db953b7b901962c8637c6f95e75ba3ca5bf8e6ae51
+size 3122285
diff --git a/taco_analysis/plots/mesh_overlay/emptyboxplate_20231027_157_f0000.png b/taco_analysis/plots/mesh_overlay/emptyboxplate_20231027_157_f0000.png
new file mode 100644
index 0000000000000000000000000000000000000000..86877d8e25d71ad43e6075fe5c87ab0f79b03405
--- /dev/null
+++ b/taco_analysis/plots/mesh_overlay/emptyboxplate_20231027_157_f0000.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a60566f1182eee6dda5b95060a4fff8d2a56c7dcc73a0f263838494324674748
+size 2806461
diff --git a/taco_analysis/plots/mesh_overlay/emptyboxplate_20231027_157_f0154.png b/taco_analysis/plots/mesh_overlay/emptyboxplate_20231027_157_f0154.png
new file mode 100644
index 0000000000000000000000000000000000000000..8d4cb2c95e151c4d1869f8041b45964dbd350718
--- /dev/null
+++ b/taco_analysis/plots/mesh_overlay/emptyboxplate_20231027_157_f0154.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9b7bdcd4274dffc7e27f0d3d51ed73bdadf75db96908b1a9b20379f6d7c6a336
+size 3223981
diff --git a/taco_analysis/plots/mesh_overlay/measurerulercup_20231104_108_f0000.png b/taco_analysis/plots/mesh_overlay/measurerulercup_20231104_108_f0000.png
new file mode 100644
index 0000000000000000000000000000000000000000..c328bd4d8f7772af7b37fdb46de03286792ff4f6
--- /dev/null
+++ b/taco_analysis/plots/mesh_overlay/measurerulercup_20231104_108_f0000.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a736b14c75ada08b6373baae66d6823afd330666dad3fc164434ac2c6633345f
+size 2557904
diff --git a/taco_analysis/plots/mesh_overlay/measurerulercup_20231104_108_f0119.png b/taco_analysis/plots/mesh_overlay/measurerulercup_20231104_108_f0119.png
new file mode 100644
index 0000000000000000000000000000000000000000..e3e3114b73ea930220f000ce78309ae34642d19d
--- /dev/null
+++ b/taco_analysis/plots/mesh_overlay/measurerulercup_20231104_108_f0119.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ead4a696fbe03a806558cbfced9e9265d05c30ddeb4f63b8cf9259ae709dab40
+size 3022535
diff --git a/taco_analysis/plots/mesh_overlay/pourinsomekettleplate_20231031_075_f0000.png b/taco_analysis/plots/mesh_overlay/pourinsomekettleplate_20231031_075_f0000.png
new file mode 100644
index 0000000000000000000000000000000000000000..12372637b72ff2fe3a04305f93716a210ddbac78
--- /dev/null
+++ b/taco_analysis/plots/mesh_overlay/pourinsomekettleplate_20231031_075_f0000.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:441d49d73d55b95ea65c24e743c51044929744af65de99fe63cd4ccc6f5d08b2
+size 2944981
diff --git a/taco_analysis/plots/mesh_overlay/pourinsomekettleplate_20231031_075_f0152.png b/taco_analysis/plots/mesh_overlay/pourinsomekettleplate_20231031_075_f0152.png
new file mode 100644
index 0000000000000000000000000000000000000000..b2dd7cdb961badfdaa0874551cf2e62251bffd91
--- /dev/null
+++ b/taco_analysis/plots/mesh_overlay/pourinsomekettleplate_20231031_075_f0152.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:44b3ad9ef4bafdd31036341ff757418b2f6ac1563baee1d32f7a062d81da9f24
+size 3451481
diff --git a/taco_analysis/plots/mesh_overlay/putoutbowlbowl_20231013_136_f0000.png b/taco_analysis/plots/mesh_overlay/putoutbowlbowl_20231013_136_f0000.png
new file mode 100644
index 0000000000000000000000000000000000000000..a08ac8b3967fbac9a6b4cf9ae21e6aa28fb7d937
--- /dev/null
+++ b/taco_analysis/plots/mesh_overlay/putoutbowlbowl_20231013_136_f0000.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f065e280b821660328900300d58826f1e1b335c916f0360b4adc8220aadf24b5
+size 2662514
diff --git a/taco_analysis/plots/mesh_overlay/putoutbowlbowl_20231013_136_f0116.png b/taco_analysis/plots/mesh_overlay/putoutbowlbowl_20231013_136_f0116.png
new file mode 100644
index 0000000000000000000000000000000000000000..8905993762f9e0f2c63f7afbcf9c3ddb5fd29c46
--- /dev/null
+++ b/taco_analysis/plots/mesh_overlay/putoutbowlbowl_20231013_136_f0116.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dfa424b541827f99a09fe8179e19c17590dec85c2b0d65069b8bb55fa2fcbed1
+size 3181190
diff --git a/taco_analysis/plots/mesh_overlay/scrapeoffknifeplate_20231020_200_f0000.png b/taco_analysis/plots/mesh_overlay/scrapeoffknifeplate_20231020_200_f0000.png
new file mode 100644
index 0000000000000000000000000000000000000000..47749fb3d2606e969be5ff42e7083a900221a135
--- /dev/null
+++ b/taco_analysis/plots/mesh_overlay/scrapeoffknifeplate_20231020_200_f0000.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e90a4e88490ac95623e0aeff98dbeda7ad0e1a58a43a558041c7641abed88058
+size 2629023
diff --git a/taco_analysis/plots/mesh_overlay/scrapeoffknifeplate_20231020_200_f0174.png b/taco_analysis/plots/mesh_overlay/scrapeoffknifeplate_20231020_200_f0174.png
new file mode 100644
index 0000000000000000000000000000000000000000..18fd7116a3f1483f651075dc4e8bc21c73c87482
--- /dev/null
+++ b/taco_analysis/plots/mesh_overlay/scrapeoffknifeplate_20231020_200_f0174.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:30b653254547d548d25fc9573ab6c8a622ad0f44d871ac01e0bb6e8dbb71b639
+size 3037697
diff --git a/taco_analysis/plots/mesh_overlay/skimoffbowlpan_20230926_017_f0000.png b/taco_analysis/plots/mesh_overlay/skimoffbowlpan_20230926_017_f0000.png
new file mode 100644
index 0000000000000000000000000000000000000000..f8fe3bac458fc8d1cc3ceae0ace8d8f3aeaf8489
--- /dev/null
+++ b/taco_analysis/plots/mesh_overlay/skimoffbowlpan_20230926_017_f0000.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3dd087371b71dc378c3581ae44cba6ceb912ad1b734918f5bf91061e280483bf
+size 2913752
diff --git a/taco_analysis/plots/mesh_overlay/skimoffbowlpan_20230926_017_f0249.png b/taco_analysis/plots/mesh_overlay/skimoffbowlpan_20230926_017_f0249.png
new file mode 100644
index 0000000000000000000000000000000000000000..8bfd7b914dd4e3abbae9c8ea1f527d9163635074
--- /dev/null
+++ b/taco_analysis/plots/mesh_overlay/skimoffbowlpan_20230926_017_f0249.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5ea6d3febeb2ced48bf65e3e3a0c4155afcc6c714b592c20b310017ce5c2bdb9
+size 3360525
diff --git a/taco_analysis/plots/mesh_overlay/stir-fryspoonpan_20231020_050_f0000.png b/taco_analysis/plots/mesh_overlay/stir-fryspoonpan_20231020_050_f0000.png
new file mode 100644
index 0000000000000000000000000000000000000000..eb3760ad5ceca47031ef8c9eefffdb782a15e9c6
--- /dev/null
+++ b/taco_analysis/plots/mesh_overlay/stir-fryspoonpan_20231020_050_f0000.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8c49a127938429890302baebd10d6c3dc3f308474eb1a1f6429cd525cb4cdb99
+size 2681707
diff --git a/taco_analysis/plots/mesh_overlay/stir-fryspoonpan_20231020_050_f0155.png b/taco_analysis/plots/mesh_overlay/stir-fryspoonpan_20231020_050_f0155.png
new file mode 100644
index 0000000000000000000000000000000000000000..cb964569a9ac65132e5f18a207cd9c97335ab715
--- /dev/null
+++ b/taco_analysis/plots/mesh_overlay/stir-fryspoonpan_20231020_050_f0155.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:496a422cb8f4788cd11cdd3d27b9c0ecf992b4a18d02989151fc4a0bd5808c38
+size 3166959
diff --git a/taco_analysis/plots/mesh_overlay/stirspatulapan_20231020_037_f0000.png b/taco_analysis/plots/mesh_overlay/stirspatulapan_20231020_037_f0000.png
new file mode 100644
index 0000000000000000000000000000000000000000..28330d22b52f11221c3c0521c361b9d8e4ba5139
--- /dev/null
+++ b/taco_analysis/plots/mesh_overlay/stirspatulapan_20231020_037_f0000.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2dc87d32b05516a0bc055ff5ea5241584d5ea6cee30af96b168aaf24beff86d6
+size 2495810
diff --git a/taco_analysis/plots/mesh_overlay/stirspatulapan_20231020_037_f0130.png b/taco_analysis/plots/mesh_overlay/stirspatulapan_20231020_037_f0130.png
new file mode 100644
index 0000000000000000000000000000000000000000..9ba3f469daf407926ccbfe9cd198c8f0daa761be
--- /dev/null
+++ b/taco_analysis/plots/mesh_overlay/stirspatulapan_20231020_037_f0130.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3014490b490e11a38a32ba0a1159ac3c743d9ccb1e5fe2598d076125f04a1bcc
+size 2946169
diff --git a/taco_analysis/plots/modality_completeness.png b/taco_analysis/plots/modality_completeness.png
new file mode 100644
index 0000000000000000000000000000000000000000..c76aa274778843c3cfbf4fcd6b802cf0bf8232b3
--- /dev/null
+++ b/taco_analysis/plots/modality_completeness.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1fcb85806661aaa8c1bd62118756f8bb2febd8ca1e8feefcc9146f987d6ea043
+size 90765
diff --git a/taco_analysis/plots/object_distribution.png b/taco_analysis/plots/object_distribution.png
new file mode 100644
index 0000000000000000000000000000000000000000..5907b9c71b413322ed1abca3ff10d648b0b4534c
--- /dev/null
+++ b/taco_analysis/plots/object_distribution.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e35a91e8d5255557fba7ffb9df20b48be1d498a3f79d434184c62a3251a115d3
+size 50742
diff --git a/taco_analysis/plots/scene_2d/brusheraserkettle_20231103_110.png b/taco_analysis/plots/scene_2d/brusheraserkettle_20231103_110.png
new file mode 100644
index 0000000000000000000000000000000000000000..42a73034c239c21811ae073c4bc8ac9652f2ab69
--- /dev/null
+++ b/taco_analysis/plots/scene_2d/brusheraserkettle_20231103_110.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ce563826344d550a40779a8a8701117f2c408acefbee7511b7da82f2c1f7ab97
+size 131994
diff --git a/taco_analysis/plots/scene_2d/measurerulerbowl_20231104_088.png b/taco_analysis/plots/scene_2d/measurerulerbowl_20231104_088.png
new file mode 100644
index 0000000000000000000000000000000000000000..d243273dcb4e3d624a69c9b770a8348bfaa135b2
--- /dev/null
+++ b/taco_analysis/plots/scene_2d/measurerulerbowl_20231104_088.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cc51a2da59fdb0e13cd42c3d39517aa198a0ba79c1790cd50e49e5ccb6b5e50d
+size 131720
diff --git a/taco_analysis/plots/scene_2d/scrapeoffknifeplate_20231015_210.png b/taco_analysis/plots/scene_2d/scrapeoffknifeplate_20231015_210.png
new file mode 100644
index 0000000000000000000000000000000000000000..216a576efe1e16b2ed49bc80c2f5802e6236f12f
--- /dev/null
+++ b/taco_analysis/plots/scene_2d/scrapeoffknifeplate_20231015_210.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4c1dfba0bf34df2cbf0b8cf1e9050df53acc0f463f491091ded1416da41baf68
+size 132809
diff --git a/taco_analysis/plots/scene_2d/skimoffspoonplate_20231024_069.png b/taco_analysis/plots/scene_2d/skimoffspoonplate_20231024_069.png
new file mode 100644
index 0000000000000000000000000000000000000000..9aef13487d6ba2639cba8f55ed0d9cf379a7e004
--- /dev/null
+++ b/taco_analysis/plots/scene_2d/skimoffspoonplate_20231024_069.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:75da0773df72907f9fe6381f8e0488353d309ca5014854143e4516821b65813e
+size 134134
diff --git a/taco_analysis/plots/scene_2d/stirspoonpan_20231020_079.png b/taco_analysis/plots/scene_2d/stirspoonpan_20231020_079.png
new file mode 100644
index 0000000000000000000000000000000000000000..c35a88ba93da7dd93f75126fc3c34d526ae85e74
--- /dev/null
+++ b/taco_analysis/plots/scene_2d/stirspoonpan_20231020_079.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6a9b151cecb3ea67e8dbb370a24cf15971e61c4fda3d995eb3c3fb7831428717
+size 130528
diff --git a/taco_analysis/plots/scene_3d/brusheraserkettle_20231103_110_view1.png b/taco_analysis/plots/scene_3d/brusheraserkettle_20231103_110_view1.png
new file mode 100644
index 0000000000000000000000000000000000000000..ea24c34d80feb3cc2ed9a37acf56b43715f9fc5c
--- /dev/null
+++ b/taco_analysis/plots/scene_3d/brusheraserkettle_20231103_110_view1.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9f902995bf95ba768069487d565dc788483b569ad0f60715fb6eefb072f6edf3
+size 300031
diff --git a/taco_analysis/plots/scene_3d/brusheraserkettle_20231103_110_view2.png b/taco_analysis/plots/scene_3d/brusheraserkettle_20231103_110_view2.png
new file mode 100644
index 0000000000000000000000000000000000000000..2c100b5d674183647d59014775463d1708c25298
--- /dev/null
+++ b/taco_analysis/plots/scene_3d/brusheraserkettle_20231103_110_view2.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ff5552048b71cd113bf983ec1abde0fe65465231aa00c4a7037f49b9751663ce
+size 368115
diff --git a/taco_analysis/plots/scene_3d/measurerulerbowl_20231104_088_view1.png b/taco_analysis/plots/scene_3d/measurerulerbowl_20231104_088_view1.png
new file mode 100644
index 0000000000000000000000000000000000000000..a09f8924c90b549984807f3b7c60f52fb544508d
--- /dev/null
+++ b/taco_analysis/plots/scene_3d/measurerulerbowl_20231104_088_view1.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7869b04272cdae693ea15f8eb1084968a4457519185c1bcbca9b771f94fb8d93
+size 298035
diff --git a/taco_analysis/plots/scene_3d/measurerulerbowl_20231104_088_view2.png b/taco_analysis/plots/scene_3d/measurerulerbowl_20231104_088_view2.png
new file mode 100644
index 0000000000000000000000000000000000000000..1da23280984ed543fc92fa4af9a8dddb8f49e8f8
--- /dev/null
+++ b/taco_analysis/plots/scene_3d/measurerulerbowl_20231104_088_view2.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3093c630f9e07809540f5ed177206398719b015becf2c9fde6beb54f8b0c30fd
+size 366919
diff --git a/taco_analysis/plots/scene_3d/scrapeoffknifeplate_20231015_210_view1.png b/taco_analysis/plots/scene_3d/scrapeoffknifeplate_20231015_210_view1.png
new file mode 100644
index 0000000000000000000000000000000000000000..973314058585b7d7509826d4f1ae13ed826a479d
--- /dev/null
+++ b/taco_analysis/plots/scene_3d/scrapeoffknifeplate_20231015_210_view1.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c13a5c62e97b4c96323a1edd94abce255a15d32dfbd547bf1bab7fa81919fb28
+size 300104
diff --git a/taco_analysis/plots/scene_3d/scrapeoffknifeplate_20231015_210_view2.png b/taco_analysis/plots/scene_3d/scrapeoffknifeplate_20231015_210_view2.png
new file mode 100644
index 0000000000000000000000000000000000000000..d579b8c24f80da97d2ffcae2f7cca9aff2bcdc23
--- /dev/null
+++ b/taco_analysis/plots/scene_3d/scrapeoffknifeplate_20231015_210_view2.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3cd39b05474e59f74507e6951582b6716bc9938ede1d7f6f58918764008e749b
+size 370293
diff --git a/taco_analysis/plots/scene_3d/skimoffspoonplate_20231024_069_view1.png b/taco_analysis/plots/scene_3d/skimoffspoonplate_20231024_069_view1.png
new file mode 100644
index 0000000000000000000000000000000000000000..cc3e19d2e683445f656e3965a13fb6f3602ef3c1
--- /dev/null
+++ b/taco_analysis/plots/scene_3d/skimoffspoonplate_20231024_069_view1.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5eb1212a3f6264465698c96264c949086ba41edf33305df67511b87fa3d08026
+size 299066
diff --git a/taco_analysis/plots/scene_3d/skimoffspoonplate_20231024_069_view2.png b/taco_analysis/plots/scene_3d/skimoffspoonplate_20231024_069_view2.png
new file mode 100644
index 0000000000000000000000000000000000000000..2554b6982b7cee0240d57fa72f0ac7798041b3bc
--- /dev/null
+++ b/taco_analysis/plots/scene_3d/skimoffspoonplate_20231024_069_view2.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b2bfd9625665f7589bdf423443e6fdcaea053dbf9d77315c305564f055cbc0ca
+size 369963
diff --git a/taco_analysis/plots/scene_3d/stirspoonpan_20231020_079_view1.png b/taco_analysis/plots/scene_3d/stirspoonpan_20231020_079_view1.png
new file mode 100644
index 0000000000000000000000000000000000000000..6f62242a0890e02b68583d78f14e976cf7aa7aca
--- /dev/null
+++ b/taco_analysis/plots/scene_3d/stirspoonpan_20231020_079_view1.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:101df0659007c9e82f1cecd54ec72be0868c232e6453f5e6042a06eaec31758a
+size 298947
diff --git a/taco_analysis/plots/scene_3d/stirspoonpan_20231020_079_view2.png b/taco_analysis/plots/scene_3d/stirspoonpan_20231020_079_view2.png
new file mode 100644
index 0000000000000000000000000000000000000000..f0aaf810c2918300539f9fadd7164ccc519dca11
--- /dev/null
+++ b/taco_analysis/plots/scene_3d/stirspoonpan_20231020_079_view2.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:543c785a3402f207048bc64f8d7e949da9ec39cb1267bb5fef01b5dca68a58e4
+size 369629
diff --git a/taco_analysis/plots/sequences_per_triplet.png b/taco_analysis/plots/sequences_per_triplet.png
new file mode 100644
index 0000000000000000000000000000000000000000..6b0abc6ff22f91c363526521f36854cedd3c5696
--- /dev/null
+++ b/taco_analysis/plots/sequences_per_triplet.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1e22c43df4b2dec080a242db437481553afeca18a90647586f4b4da39a3b4a2f
+size 49320
diff --git a/taco_analysis/plots/tool_distribution.png b/taco_analysis/plots/tool_distribution.png
new file mode 100644
index 0000000000000000000000000000000000000000..a7a133768f457be99f55df99718fdb45919f4535
--- /dev/null
+++ b/taco_analysis/plots/tool_distribution.png
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ffdadc11c17c19df087049b65c6f05a7e51d008690aedefef7c4a697467f62b0
+size 73865
diff --git a/taco_analysis/render_mesh_overlay.py b/taco_analysis/render_mesh_overlay.py
new file mode 100644
index 0000000000000000000000000000000000000000..19e0b8f033b9da65e46400c079f489545a2ca2af
--- /dev/null
+++ b/taco_analysis/render_mesh_overlay.py
@@ -0,0 +1,549 @@
+#!/usr/bin/env python
+"""Render object meshes + hand skeletons onto all 12 allocentric views.
+
+For each selected sequence + frame, projects:
+  - Tool mesh (blue) and target mesh (coral) with filled triangles + wireframe
+  - Left hand skeleton (green) and right hand skeleton (yellow)
+
+Camera convention:  p_cam = R @ p_world + T   (world-to-camera)
+
+Hand poses: hand_pose (48,) = [0:3] global wrist rotation (axis-angle)
+  + [3:48] 15 finger joint rotations (axis-angle, NOT PCA).
+  hand_trans (3,) = wrist position in world (meters).
+  hand_shape (10,) = MANO beta parameters (optional).
+  Joints are computed via MANO forward kinematics (LBS).
+
+Usage:
+    python render_mesh_overlay.py --n-sequences 5 --seed 77
+"""
+
+import argparse
+import inspect
+import json
+import pickle
+from pathlib import Path
+
+# Monkey-patch for chumpy compat with Python 3.12+ / numpy 2.x
+if not hasattr(inspect, "getargspec"):
+    inspect.getargspec = inspect.getfullargspec
+import numpy as np
+for _alias in ("bool", "int", "float", "complex", "object", "str"):
+    if not hasattr(np, _alias):
+        setattr(np, _alias, getattr(__builtins__, _alias, object))
+if not hasattr(np, "unicode"):
+    np.unicode = str
+
+import cv2
+import imageio
+import pandas as pd
+import scipy.sparse
+import trimesh
+from PIL import Image, ImageDraw
+
+
+MESH_COLORS = {
+    "tool": (50, 120, 255),     # blue (slightly darker/more saturated)
+    "target": (255, 90, 60),    # coral (slightly darker)
+}
+
+HAND_COLORS = {
+    "left":  (50, 220, 100),    # green
+    "right": (255, 220, 50),    # yellow
+}
+
+# Bone connections for drawing (21-joint order after MANO reorder):
+# 0=wrist, 1-4=index, 5-8=middle, 9-12=ring, 13-16=pinky, 17-20=thumb
+HAND_BONES = [
+    (0, 1), (1, 2), (2, 3), (3, 4),       # Index
+    (0, 5), (5, 6), (6, 7), (7, 8),       # Middle
+    (0, 9), (9, 10), (10, 11), (11, 12),  # Ring
+    (0, 13), (13, 14), (14, 15), (15, 16), # Pinky
+    (0, 17), (17, 18), (18, 19), (19, 20), # Thumb
+    (1, 5), (5, 9), (9, 13),               # Palm cross-links
+]
+
+# MANO joint reorder: maps 16 kinematic joints + 5 fingertip vertices -> 21
+# Output order: wrist, index(MCP,PIP,DIP,tip), middle(...), ring(...), pinky(...), thumb(...)
+_JOINT_REORDER = [0, 13, 14, 15, 16, 1, 2, 3, 17, 4, 5, 6, 18, 10, 11, 12, 19, 7, 8, 9, 20]
+
+# Fingertip vertex indices in MANO mesh (thumb, index, middle, ring, pinky)
+_TIPS_RIGHT = [745, 317, 444, 556, 673]
+_TIPS_LEFT = [745, 317, 445, 556, 673]
+
+
+# ── MANO numpy forward pass ─────────────────────────────────────────
+
+def _load_mano_pkl(path):
+    """Load a MANO .pkl, converting chumpy arrays to numpy."""
+    with open(path, "rb") as f:
+        raw = pickle.load(f, encoding="latin1")
+    out = {}
+    for k, v in raw.items():
+        if hasattr(v, "r"):
+            out[k] = np.array(v.r, dtype=np.float64)
+        elif scipy.sparse.issparse(v):
+            out[k] = v
+        elif isinstance(v, np.ndarray):
+            out[k] = v.astype(np.float64) if v.dtype.kind == "f" else v
+        else:
+            out[k] = v
+    return out
+
+
+def _rodrigues(rotvec):
+    """Axis-angle (3,) -> rotation matrix (3,3)."""
+    angle = np.linalg.norm(rotvec)
+    if angle < 1e-8:
+        return np.eye(3, dtype=np.float64)
+    k = rotvec / angle
+    K = np.array([[0, -k[2], k[1]],
+                  [k[2], 0, -k[0]],
+                  [-k[1], k[0], 0]], dtype=np.float64)
+    return np.eye(3) + np.sin(angle) * K + (1 - np.cos(angle)) * (K @ K)
+
+
+def _batch_rodrigues(axisang):
+    """(N, 3) axis-angle -> (N, 3, 3) rotation matrices."""
+    return np.array([_rodrigues(a) for a in axisang])
+
+
+class ManoModel:
+    """Numpy-only MANO forward kinematics (no PyTorch/GPU needed).
+
+    Mirrors manopth ManoLayer with use_pca=False, flat_hand_mean=True, center_idx=0.
+    """
+
+    def __init__(self, mano_pkl_path, side="right"):
+        d = _load_mano_pkl(mano_pkl_path)
+        self.v_template = d["v_template"]              # (778, 3)
+        self.shapedirs = d["shapedirs"]                # (778, 3, 10)
+        self.posedirs = d["posedirs"]                  # (778, 3, 135)
+        self.weights = d["weights"]                    # (778, 16)
+        self.faces = d["f"]                            # (F, 3)
+        self.side = side
+
+        jr = d["J_regressor"]
+        self.J_regressor = jr.toarray() if scipy.sparse.issparse(jr) else np.array(jr)
+
+        kt = d["kintree_table"]
+        self.parents = [int(kt[0, i]) for i in range(kt.shape[1])]
+        self.parents[0] = -1  # root has no parent
+
+        self.tip_indices = _TIPS_RIGHT if side == "right" else _TIPS_LEFT
+
+    def forward(self, hand_pose, hand_trans, betas=None):
+        """Compute 21 joint positions in world coordinates (meters).
+
+        Args:
+            hand_pose: (48,) = [0:3] global rot + [3:48] 15 joint axis-angles
+            hand_trans: (3,) wrist translation in meters
+            betas: (10,) shape params or None (uses mean shape)
+
+        Returns:
+            joints: (21, 3) in meters, world frame
+        """
+        # 1. Shape blend shapes
+        if betas is not None and self.shapedirs.ndim == 3:
+            v_shaped = self.v_template + np.einsum("vci,i->vc", self.shapedirs, betas)
+        else:
+            v_shaped = self.v_template.copy()
+
+        # 2. Joint locations from shaped mesh
+        J = self.J_regressor @ v_shaped  # (16, 3)
+
+        # 3. Full pose: 16 rotations (1 global + 15 finger)
+        full_pose = hand_pose[:48].copy()  # use_pca=False, flat_hand_mean -> identity
+        rot_mats = _batch_rodrigues(full_pose.reshape(16, 3))  # (16, 3, 3)
+
+        # 4. Pose blend shapes (based on rotations 1..15, excluding global)
+        pose_feature = (rot_mats[1:] - np.eye(3)).ravel()  # (135,)
+        v_posed = v_shaped + np.einsum("vci,i->vc", self.posedirs, pose_feature)
+
+        # 5. Forward kinematics
+        # Build 4x4 transforms for each joint
+        G = np.zeros((16, 4, 4), dtype=np.float64)
+
+        # Root
+        G[0, :3, :3] = rot_mats[0]
+        G[0, :3, 3] = J[0]
+        G[0, 3, 3] = 1.0
+
+        for i in range(1, 16):
+            p = self.parents[i]
+            local = np.eye(4, dtype=np.float64)
+            local[:3, :3] = rot_mats[i]
+            local[:3, 3] = J[i] - J[p]
+            G[i] = G[p] @ local
+
+        # 6. Joint positions from global transforms (BEFORE rest-pose correction)
+        jtr = G[:, :3, 3].copy()  # (16, 3)
+
+        # Remove rest-pose joint offsets (only needed for LBS skinning)
+        G_skin = G.copy()
+        for i in range(16):
+            Jh = np.append(J[i], 0.0)  # (4,)
+            G_skin[i, :, 3] -= G_skin[i] @ Jh
+
+        # 7. LBS skinning (for fingertip vertex positions only)
+        T = np.einsum("jab,vj->vab", G_skin, self.weights)  # (778, 4, 4)
+        v_homo = np.concatenate([v_posed, np.ones((778, 1))], axis=1)  # (778, 4)
+        v_final = np.einsum("vab,vb->va", T, v_homo)[:, :3]  # (778, 3)
+
+        # 8. Fingertip positions from skinned vertices
+        tips = v_final[self.tip_indices]  # (5, 3)
+        jtr = np.concatenate([jtr, tips], axis=0)  # (21, 3)
+
+        # 9. Center on wrist (center_idx=0), then add translation
+        # MANO template is in meters; manolayer *1000 and loader /1000 cancel out.
+        center = jtr[0:1].copy()
+        jtr -= center
+
+        # 10. Translate to world
+        jtr += hand_trans
+
+        # 12. Reorder to visualization order
+        return jtr[_JOINT_REORDER]
+
+
+# Module-level cache for MANO models
+_mano_models = {}
+
+
+def _get_mano_model(mano_root, side):
+    """Get or load a cached ManoModel."""
+    key = (str(mano_root), side)
+    if key not in _mano_models:
+        fname = "MANO_LEFT.pkl" if side == "left" else "MANO_RIGHT.pkl"
+        _mano_models[key] = ManoModel(mano_root / fname, side=side)
+    return _mano_models[key]
+
+
+# ── Core rendering ────────────────────────────────────────────────────
+
+def load_calibration(path):
+    with open(path) as f:
+        return json.load(f)
+
+
+def read_frame(video_path, frame_idx):
+    reader = imageio.get_reader(str(video_path))
+    frame = reader.get_data(frame_idx)
+    reader.close()
+    return frame
+
+
+def project_points(pts_world, K, R, T):
+    """Project Nx3 world points to Nx2 pixel coords + Nx depths."""
+    p_cam = (R @ pts_world.T).T + T
+    depth = p_cam[:, 2].copy()
+    safe_z = np.where(np.abs(depth) > 1e-6, depth, 1e-6)
+    u = K[0, 0] * p_cam[:, 0] / safe_z + K[0, 2]
+    v = K[1, 1] * p_cam[:, 1] / safe_z + K[1, 2]
+    return np.stack([u, v], axis=1), depth
+
+
+def render_mesh_on_image(img, verts_world, faces, K, R, T,
+                         color=(70, 140, 255), fill_alpha=1.0,
+                         edge_color=None, max_faces=5000):
+    """Render a mesh overlay: filled triangles + wireframe edges."""
+    H, W = img.shape[:2]
+    uv, depth = project_points(verts_world, K, R, T)
+
+    face_depths = depth[faces]
+    valid_mask = np.all(face_depths > 0.01, axis=1)
+    valid_faces = faces[valid_mask]
+    if len(valid_faces) == 0:
+        return img
+
+    if len(valid_faces) > max_faces:
+        idx = np.random.choice(len(valid_faces), max_faces, replace=False)
+        valid_faces = valid_faces[idx]
+
+    order = np.argsort(-depth[valid_faces].mean(axis=1))
+    valid_faces = valid_faces[order]
+    uv_int = uv.astype(np.int32)
+
+    overlay = img.copy()
+    for face in valid_faces:
+        pts = uv_int[face]
+        if (pts[:, 0].max() < -50 or pts[:, 0].min() > W + 50 or
+            pts[:, 1].max() < -50 or pts[:, 1].min() > H + 50):
+            continue
+        cv2.fillPoly(overlay, [pts.reshape(-1, 1, 2)], color)
+    result = cv2.addWeighted(overlay, fill_alpha, img, 1.0 - fill_alpha, 0)
+
+    ec = edge_color or tuple(min(255, c + 30) for c in color)
+    for face in valid_faces[::2]:
+        pts = uv_int[face]
+        for j in range(3):
+            cv2.line(result, tuple(pts[j]), tuple(pts[(j + 1) % 3]), ec, 1, cv2.LINE_AA)
+
+    return result
+
+
+def draw_hand_skeleton(img, joints_world, K, R, T, color, scale_x, scale_y):
+    """Draw a hand skeleton (21 joints + bones) projected onto the image."""
+    uv, depth = project_points(joints_world, K, R, T)
+    H, W = img.shape[:2]
+
+    uv_s = uv.copy()
+    uv_s[:, 0] *= scale_x
+    uv_s[:, 1] *= scale_y
+    uv_int = uv_s.astype(np.int32)
+
+    joint_color = tuple(min(255, c + 60) for c in color)
+    outline_color = tuple(max(0, c - 80) for c in color)
+
+    bone_w = max(2, int(min(W, H) / 200))
+    joint_r = max(3, int(min(W, H) / 150))
+
+    # Draw bones
+    for j1, j2 in HAND_BONES:
+        if depth[j1] < 0.01 or depth[j2] < 0.01:
+            continue
+        p1, p2 = tuple(uv_int[j1]), tuple(uv_int[j2])
+        if (p1[0] < -50 and p2[0] < -50) or (p1[0] > W + 50 and p2[0] > W + 50):
+            continue
+        if (p1[1] < -50 and p2[1] < -50) or (p1[1] > H + 50 and p2[1] > H + 50):
+            continue
+        cv2.line(img, p1, p2, outline_color, bone_w + 2, cv2.LINE_AA)
+        cv2.line(img, p1, p2, joint_color, bone_w, cv2.LINE_AA)
+
+    # Draw joints
+    for i in range(len(joints_world)):
+        if depth[i] < 0.01:
+            continue
+        u_s, v_s = uv_int[i]
+        if not (-30 <= u_s < W + 30 and -30 <= v_s < H + 30):
+            continue
+        # Tips slightly smaller, MCP/wrist larger
+        r = joint_r + 1 if i in (0, 1, 5, 9, 13, 17) else joint_r
+        cv2.circle(img, (u_s, v_s), r, outline_color, -1, cv2.LINE_AA)
+        cv2.circle(img, (u_s, v_s), max(1, r - 1), joint_color, -1, cv2.LINE_AA)
+        if i == 0:  # Wrist emphasis
+            cv2.circle(img, (u_s, v_s), r + 2, (255, 255, 255), 1, cv2.LINE_AA)
+
+    return img
+
+
+# ── Hand data loading ─────────────────────────────────────────────────
+
+def load_hand_skeletons(root, row, frame_idx, mano_root=None):
+    """Load hand skeleton joint positions via MANO forward kinematics.
+
+    Returns dict: {'left': (21, 3), 'right': (21, 3)} or empty if missing.
+    """
+    hand_dir = row.get("hand_poses_dir", "")
+    if not hand_dir or pd.isna(hand_dir):
+        return {}
+
+    if mano_root is None:
+        mano_root = root.parent / "mano_v1_2" / "models"
+
+    result = {}
+    hand_path = root / hand_dir
+    frame_key = f"{frame_idx + 1:05d}"
+
+    for side in ["left", "right"]:
+        pose_file = hand_path / f"{side}_hand.pkl"
+        if not pose_file.exists():
+            continue
+        with open(pose_file, "rb") as f:
+            data = pickle.load(f)
+        if frame_key not in data:
+            continue
+        frame_data = data[frame_key]
+        if not isinstance(frame_data, dict):
+            continue
+        if "hand_pose" not in frame_data or "hand_trans" not in frame_data:
+            continue
+
+        hand_pose = np.array(frame_data["hand_pose"], dtype=np.float64).ravel()
+        hand_trans = np.array(frame_data["hand_trans"], dtype=np.float64).ravel()
+
+        # Load shape parameters (betas) if available
+        betas = None
+        shape_file = hand_path / f"{side}_hand_shape.pkl"
+        if shape_file.exists():
+            with open(shape_file, "rb") as f:
+                shape_data = pickle.load(f)
+            betas = np.array(shape_data["hand_shape"], dtype=np.float64).ravel()[:10]
+
+        mano = _get_mano_model(mano_root, side)
+        joints = mano.forward(hand_pose, hand_trans, betas=betas)
+        result[side] = joints
+
+    return result
+
+
+# ── Grid builder ──────────────────────────────────────────────────────
+
+def build_grid(root, row, calib, frame_idx, thumb_width=640, mano_root=None):
+    """Build a 4x3 grid of mesh + hand skeleton overlaid camera views."""
+
+    camera_ids = row["camera_ids"].split(";")
+    mr_dir = root / row["marker_removed_dir"]
+    models_dir = root / "object_models_released"
+
+    # --- Load object meshes ---
+    meshes = []
+    obj_dir_str = row.get("object_poses_dir", "")
+    if obj_dir_str and pd.notna(obj_dir_str):
+        obj_path = root / obj_dir_str
+        for npy_file in sorted(obj_path.glob("*.npy")):
+            poses = np.load(str(npy_file))
+            if frame_idx >= len(poses):
+                continue
+            pose = poses[frame_idx].astype(np.float64)
+            parts = npy_file.stem.split("_")
+            role, obj_id = parts[0], parts[1]
+            model_path = models_dir / f"{obj_id}_cm.obj"
+            if not model_path.exists():
+                continue
+            mesh = trimesh.load(str(model_path), process=False)
+            verts = np.array(mesh.vertices, dtype=np.float64) / 100.0
+            verts_world = (pose[:3, :3] @ verts.T).T + pose[:3, 3]
+            color = MESH_COLORS.get(role, (180, 180, 180))
+            meshes.append((verts_world, np.array(mesh.faces), color, f"{role}_{obj_id}"))
+
+    # --- Load hand skeletons ---
+    hands = load_hand_skeletons(root, row, frame_idx, mano_root=mano_root)
+
+    # --- Render each camera tile ---
+    label_h = 26
+    tiles = []
+
+    for cam_id in camera_ids:
+        frame = read_frame(mr_dir / f"{cam_id}.mp4", frame_idx)
+        H_orig, W_orig = frame.shape[:2]
+
+        cam_data = calib[cam_id]
+        K = np.array(cam_data["K"], dtype=np.float64).reshape(3, 3)
+        R = np.array(cam_data["R"], dtype=np.float64).reshape(3, 3)
+        T = np.array(cam_data["T"], dtype=np.float64).reshape(3)
+
+        # Render object meshes (on full-res image)
+        img = frame.copy()
+        for verts_world, faces_arr, color, label in meshes:
+            img = render_mesh_on_image(img, verts_world, faces_arr, K, R, T,
+                                       color=color, fill_alpha=1.0, max_faces=5000)
+
+        # Resize to thumbnail
+        thumb_h = int(thumb_width * H_orig / W_orig)
+        scale_x = thumb_width / W_orig
+        scale_y = thumb_h / H_orig
+        img_resized = cv2.resize(img, (thumb_width, thumb_h), interpolation=cv2.INTER_AREA)
+
+        # Draw hand skeletons (on thumbnail coords)
+        for side, joints in hands.items():
+            hand_color = HAND_COLORS[side]
+            img_resized = draw_hand_skeleton(
+                img_resized, joints, K, R, T,
+                hand_color, scale_x, scale_y,
+            )
+
+        # Assemble tile with label
+        tile = Image.new("RGB", (thumb_width, thumb_h + label_h), (25, 25, 25))
+        tile_draw = ImageDraw.Draw(tile)
+        tile_draw.text((4, 4), f"cam {cam_id}", fill=(200, 200, 200))
+        tile.paste(Image.fromarray(img_resized), (0, label_h))
+        tiles.append(tile)
+
+    # --- Arrange 4x3 grid ---
+    n_cols, n_rows = 4, 3
+    tile_w, tile_h = tiles[0].size
+    gap = 3
+    title_h = 40
+
+    grid_w = n_cols * tile_w + (n_cols - 1) * gap
+    grid_h = title_h + n_rows * tile_h + (n_rows - 1) * gap
+    grid = Image.new("RGB", (grid_w, grid_h), (15, 15, 15))
+    gd = ImageDraw.Draw(grid)
+
+    mesh_labels = ", ".join(lbl for _, _, _, lbl in meshes)
+    hand_labels = ", ".join(f"{s} hand" for s in hands)
+    title = f"{row['sequence_id']}  |  frame {frame_idx}  |  {mesh_labels}  |  {hand_labels}"
+    gd.text((10, 8), title, fill=(255, 255, 255))
+
+    # Legend
+    x_leg = grid_w - 450
+    for role, color in MESH_COLORS.items():
+        gd.rectangle([x_leg, 10, x_leg + 14, 24], fill=color)
+        gd.text((x_leg + 18, 8), role, fill=color)
+        x_leg += 80
+    for side, color in HAND_COLORS.items():
+        gd.ellipse([x_leg, 11, x_leg + 12, 23], fill=color)
+        gd.text((x_leg + 16, 8), f"{side[0].upper()} hand", fill=color)
+        x_leg += 90
+
+    for i, tile in enumerate(tiles):
+        r, c = divmod(i, n_cols)
+        x = c * (tile_w + gap)
+        y = title_h + r * (tile_h + gap)
+        grid.paste(tile, (x, y))
+
+    return grid
+
+
+def main():
+    parser = argparse.ArgumentParser(description="Render mesh + hand overlays on TACO views")
+    parser.add_argument("--csv", type=Path,
+                        default=Path(__file__).resolve().parent.parent / "taco_dataset" / "taco_info.csv")
+    parser.add_argument("--root", type=Path,
+                        default=Path(__file__).resolve().parent.parent / "taco_dataset")
+    parser.add_argument("--mano-root", type=Path,
+                        default=Path(__file__).resolve().parent.parent / "mano_v1_2" / "models")
+    parser.add_argument("--save-dir", type=Path,
+                        default=Path(__file__).resolve().parent / "plots" / "mesh_overlay")
+    parser.add_argument("--n-sequences", type=int, default=5)
+    parser.add_argument("--frames-per-seq", type=int, default=2)
+    parser.add_argument("--thumb-width", type=int, default=512)
+    parser.add_argument("--seed", type=int, default=77)
+    args = parser.parse_args()
+
+    df = pd.read_csv(args.csv)
+    df = df[df["all_modalities_complete"] & (df["n_allocentric_cameras"] == 12)].reset_index(drop=True)
+    args.save_dir.mkdir(parents=True, exist_ok=True)
+    root = args.root.resolve()
+
+    rng = np.random.RandomState(args.seed)
+    actions = sorted(df["action"].unique())
+    rng.shuffle(actions)
+
+    sampled = []
+    for action in actions:
+        subset = df[df["action"] == action]
+        idx = subset.index[rng.randint(0, len(subset))]
+        if idx not in sampled:
+            sampled.append(idx)
+        if len(sampled) >= args.n_sequences:
+            break
+
+    total = 0
+    for i, seq_idx in enumerate(sampled):
+        row = df.iloc[seq_idx]
+        seq_id = row["sequence_id"]
+        n_frames = int(row["n_frames"])
+
+        calib_path = root / row["alloc_camera_params_path"]
+        with open(calib_path) as f:
+            calib = json.load(f)
+
+        frame_indices = np.linspace(0, n_frames - 1, args.frames_per_seq, dtype=int).tolist()
+
+        print(f"[{i+1}/{len(sampled)}] {seq_id}  ({n_frames} frames)")
+        for fidx in frame_indices:
+            print(f"  frame {fidx}...", end=" ", flush=True)
+            grid = build_grid(root, row, calib, fidx, thumb_width=args.thumb_width,
+                              mano_root=args.mano_root)
+            safe = seq_id.replace("/", "_").replace(" ", "").replace(",", "").replace("(", "").replace(")", "")
+            fname = f"{safe}_f{fidx:04d}.png"
+            grid.save(args.save_dir / fname)
+            print(f"saved {fname}")
+            total += 1
+
+    print(f"\nDone — {total} grid images saved to {args.save_dir}/")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/taco_analysis/visualize_taco_3d_scene.py b/taco_analysis/visualize_taco_3d_scene.py
new file mode 100644
index 0000000000000000000000000000000000000000..15eb28d43cf5e45da266d5e7ded73d1d2a03e6be
--- /dev/null
+++ b/taco_analysis/visualize_taco_3d_scene.py
@@ -0,0 +1,275 @@
+#!/usr/bin/env python
+"""Visualize TACO camera setup and object meshes in 3D.
+
+For selected sequences, plots:
+  - 12 allocentric camera positions + frustums (colour-coded)
+  - Egocentric camera position + trajectory (if available)
+  - Object meshes at frame 0 pose (tool + target)
+
+Camera convention (from TACO dataset_utils):
+  calibration.json stores per-camera:
+    K: (9,) flat 3x3 intrinsic (row-major)
+    R: (9,) flat 3x3 rotation (world-to-camera)
+    T: (3,) translation (world-to-camera)
+  So: p_cam = R @ p_world + T
+  Camera center in world = -R^T @ T
+
+Usage:
+    python visualize_taco_3d_scene.py \
+        --csv ../taco_dataset/taco_info.csv \
+        --root ../taco_dataset \
+        --save-dir plots/scene_3d
+"""
+
+import argparse
+import json
+from pathlib import Path
+
+import matplotlib
+matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+import numpy as np
+import pandas as pd
+import trimesh
+from mpl_toolkits.mplot3d.art3d import Poly3DCollection
+
+
+# ── Camera geometry helpers ──────────────────────────────────────────
+
+def camera_center_world(R, T):
+    """Camera center in world coords: c = -R^T @ T."""
+    return -R.T @ T
+
+
+def camera_axes_world(R, scale=0.15):
+    """Camera axis directions in world frame (right, down, forward)."""
+    # OpenCV convention: x=right, y=down, z=forward
+    right   = R.T @ np.array([1, 0, 0]) * scale
+    down    = R.T @ np.array([0, 1, 0]) * scale
+    forward = R.T @ np.array([0, 0, 1]) * scale
+    return right, down, forward
+
+
+def camera_frustum_world(R, T, K, w, h, scale=0.12):
+    """Return 5 points defining a camera frustum pyramid in world coords.
+
+    Points: [center, top-left, top-right, bottom-right, bottom-left]
+    """
+    c = camera_center_world(R, T)
+    fx, fy = K[0, 0], K[1, 1]
+    cx, cy = K[0, 2], K[1, 2]
+
+    # Four image corners in normalized camera coords, at depth=scale
+    corners_cam = np.array([
+        [(-cx) / fx * scale,     (-cy) / fy * scale,     scale],
+        [(w - cx) / fx * scale,  (-cy) / fy * scale,     scale],
+        [(w - cx) / fx * scale,  (h - cy) / fy * scale,  scale],
+        [(-cx) / fx * scale,     (h - cy) / fy * scale,  scale],
+    ])
+
+    # Transform to world
+    corners_world = (R.T @ corners_cam.T).T + c  # (4, 3)
+    return c, corners_world
+
+
+# ── Plotting ─────────────────────────────────────────────────────────
+
+def draw_frustum(ax, center, corners, color, alpha=0.15, label=None):
+    """Draw a camera frustum pyramid."""
+    # Lines from center to each corner
+    for corner in corners:
+        ax.plot3D(*zip(center, corner), color=color, linewidth=1.0, alpha=0.8)
+    # Rectangle at far plane
+    rect = list(corners) + [corners[0]]
+    ax.plot3D(*zip(*rect), color=color, linewidth=0.8, alpha=0.6)
+    # Filled faces
+    faces = [[center, corners[i], corners[(i + 1) % 4]] for i in range(4)]
+    faces.append(list(corners))
+    poly = Poly3DCollection(faces, alpha=alpha, facecolor=color, edgecolor=color, linewidth=0.3)
+    ax.add_collection3d(poly)
+    if label:
+        ax.text(*center, label, fontsize=5, color=color, ha="center")
+
+
+def draw_mesh(ax, vertices, faces, color, alpha=0.3, max_faces=2000):
+    """Draw a mesh as a semi-transparent Poly3DCollection."""
+    if len(faces) > max_faces:
+        # Subsample faces for performance
+        idx = np.random.choice(len(faces), max_faces, replace=False)
+        faces = faces[idx]
+    tris = vertices[faces]
+    poly = Poly3DCollection(tris, alpha=alpha, facecolor=color, edgecolor="none")
+    ax.add_collection3d(poly)
+
+
+def plot_scene(
+    calib: dict,
+    root: Path,
+    row: pd.Series,
+    frame_idx: int = 0,
+    title: str = "",
+    figsize: tuple = (16, 12),
+) -> plt.Figure:
+    """Create a 3D scene plot with cameras and objects."""
+
+    fig = plt.figure(figsize=figsize)
+    ax = fig.add_subplot(111, projection="3d")
+
+    # Colour palette for 12 cameras
+    cmap = plt.cm.tab20(np.linspace(0, 1, 12))
+
+    all_positions = []
+
+    # ── Allocentric cameras ──
+    for i, cam_id in enumerate(sorted(calib.keys())):
+        data = calib[cam_id]
+        R = np.array(data["R"], dtype=np.float64).reshape(3, 3)
+        T = np.array(data["T"], dtype=np.float64).reshape(3)
+        K = np.array(data["K"], dtype=np.float64).reshape(3, 3)
+        w, h = data["imgSize"]
+
+        center, corners = camera_frustum_world(R, T, K, w, h, scale=0.12)
+        all_positions.append(center)
+        draw_frustum(ax, center, corners, color=cmap[i], alpha=0.12, label=cam_id)
+
+    all_positions = np.array(all_positions)
+
+    # ── Egocentric camera trajectory ──
+    ego_cam_dir = row.get("ego_camera_params_dir", "")
+    if ego_cam_dir and pd.notna(ego_cam_dir):
+        ego_ext_path = root / ego_cam_dir / "egocentric_frame_extrinsic.npy"
+        if ego_ext_path.exists():
+            ego_exts = np.load(str(ego_ext_path))  # (N, 4, 4) world-to-camera
+            ego_centers = []
+            for ext in ego_exts:
+                R_ego = ext[:3, :3]
+                T_ego = ext[:3, 3]
+                ego_centers.append(camera_center_world(R_ego, T_ego))
+            ego_centers = np.array(ego_centers)
+            ax.plot3D(ego_centers[:, 0], ego_centers[:, 1], ego_centers[:, 2],
+                      color="magenta", linewidth=1.5, alpha=0.7, label="ego trajectory")
+            ax.scatter(*ego_centers[frame_idx], color="magenta", s=60, marker="^",
+                       zorder=5, label=f"ego frame {frame_idx}")
+
+    # ── Object meshes at frame pose ──
+    obj_dir = row.get("object_poses_dir", "")
+    obj_colors = {"tool": "dodgerblue", "target": "coral"}
+    models_dir = root / "object_models_released"
+
+    if obj_dir and pd.notna(obj_dir):
+        obj_path = root / obj_dir
+        for npy_file in sorted(obj_path.glob("*.npy")):
+            poses = np.load(str(npy_file))  # (N, 4, 4)
+            if frame_idx >= len(poses):
+                continue
+            pose = poses[frame_idx].astype(np.float64)  # object-to-world
+
+            # Extract object ID from filename (e.g., "tool_035.npy" → "035")
+            stem = npy_file.stem  # "tool_035"
+            parts = stem.split("_")
+            role = parts[0]  # "tool" or "target"
+            obj_id = parts[1]  # "035"
+            model_path = models_dir / f"{obj_id}_cm.obj"
+
+            if model_path.exists():
+                mesh = trimesh.load(str(model_path), process=False)
+                verts = np.array(mesh.vertices, dtype=np.float64)
+                # "_cm" suffix means units are centimeters → convert to meters
+                verts = verts / 100.0
+                # Apply pose (object-to-world)
+                verts_world = (pose[:3, :3] @ verts.T).T + pose[:3, 3]
+                faces_arr = np.array(mesh.faces)
+                color = obj_colors.get(role, "gray")
+                draw_mesh(ax, verts_world, faces_arr, color=color, alpha=0.25)
+                centroid = verts_world.mean(axis=0)
+                ax.text(*centroid, f"{role}_{obj_id}", fontsize=7, color=color)
+
+    # ── Scene origin axes ──
+    axis_len = 0.3
+    ax.quiver(0, 0, 0, axis_len, 0, 0, color="red", arrow_length_ratio=0.1, linewidth=2)
+    ax.quiver(0, 0, 0, 0, axis_len, 0, color="green", arrow_length_ratio=0.1, linewidth=2)
+    ax.quiver(0, 0, 0, 0, 0, axis_len, color="blue", arrow_length_ratio=0.1, linewidth=2)
+    ax.text(axis_len * 1.1, 0, 0, "X", color="red", fontsize=9)
+    ax.text(0, axis_len * 1.1, 0, "Y", color="green", fontsize=9)
+    ax.text(0, 0, axis_len * 1.1, "Z", color="blue", fontsize=9)
+
+    # ── Formatting ──
+    ax.set_xlabel("X (m)")
+    ax.set_ylabel("Y (m)")
+    ax.set_zlabel("Z (m)")
+    ax.set_title(title, fontsize=12, pad=20)
+
+    # Set equal aspect ratio
+    all_pts = all_positions.copy()
+    if len(all_pts) > 0:
+        mid = all_pts.mean(axis=0)
+        max_range = (all_pts.max(axis=0) - all_pts.min(axis=0)).max() / 2 * 1.3
+        ax.set_xlim(mid[0] - max_range, mid[0] + max_range)
+        ax.set_ylim(mid[1] - max_range, mid[1] + max_range)
+        ax.set_zlim(mid[2] - max_range, mid[2] + max_range)
+
+    ax.legend(fontsize=8, loc="upper right")
+    return fig
+
+
+# ── Main ─────────────────────────────────────────────────────────────
+
+def main():
+    parser = argparse.ArgumentParser(description="3D visualization of TACO cameras + objects")
+    parser.add_argument("--csv", type=Path,
+                        default=Path(__file__).resolve().parent.parent / "taco_dataset" / "taco_info.csv")
+    parser.add_argument("--root", type=Path,
+                        default=Path(__file__).resolve().parent.parent / "taco_dataset")
+    parser.add_argument("--save-dir", type=Path,
+                        default=Path(__file__).resolve().parent / "plots" / "scene_3d")
+    parser.add_argument("--n-sequences", type=int, default=5)
+    parser.add_argument("--seed", type=int, default=42)
+    args = parser.parse_args()
+
+    df = pd.read_csv(args.csv)
+    df = df[df["all_modalities_complete"]].reset_index(drop=True)
+    args.save_dir.mkdir(parents=True, exist_ok=True)
+    root = args.root.resolve()
+
+    rng = np.random.RandomState(args.seed)
+    actions = sorted(df["action"].unique())
+    rng.shuffle(actions)
+
+    sampled = []
+    for action in actions:
+        subset = df[df["action"] == action]
+        idx = subset.index[rng.randint(0, len(subset))]
+        sampled.append(idx)
+        if len(sampled) >= args.n_sequences:
+            break
+
+    for i, seq_idx in enumerate(sampled):
+        row = df.iloc[seq_idx]
+        seq_id = row["sequence_id"]
+        print(f"[{i+1}/{len(sampled)}] {seq_id}")
+
+        calib_path = root / row["alloc_camera_params_path"]
+        with open(calib_path) as f:
+            calib = json.load(f)
+
+        # Render from two viewpoints
+        for elev, azim, suffix in [(25, -60, "view1"), (45, 30, "view2")]:
+            fig = plot_scene(
+                calib=calib,
+                root=root,
+                row=row,
+                frame_idx=0,
+                title=f"{seq_id}  (frame 0)",
+            )
+            fig.axes[0].view_init(elev=elev, azim=azim)
+            safe_name = seq_id.replace("/", "_").replace(" ", "").replace(",", "").replace("(", "").replace(")", "")
+            fname = f"{safe_name}_{suffix}.png"
+            fig.savefig(args.save_dir / fname, dpi=150, bbox_inches="tight")
+            plt.close(fig)
+            print(f"  saved {fname}")
+
+    print(f"\nDone — {len(sampled) * 2} images saved to {args.save_dir}/")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/taco_analysis/visualize_taco_cameras_topdown.py b/taco_analysis/visualize_taco_cameras_topdown.py
new file mode 100644
index 0000000000000000000000000000000000000000..12bcf0f668e14e17e69b9d12b3935666d079976c
--- /dev/null
+++ b/taco_analysis/visualize_taco_cameras_topdown.py
@@ -0,0 +1,208 @@
+#!/usr/bin/env python
+"""Bird's-eye (XY) and side (XZ) views of TACO allocentric cameras + objects.
+
+Produces a 2-panel figure: top-down XY plane and side XZ plane.
+Cameras shown as triangles pointing in their forward direction.
+Objects shown as scatter clouds.
+
+Usage:
+    python visualize_taco_cameras_topdown.py
+"""
+
+import argparse
+import json
+from pathlib import Path
+
+import matplotlib
+matplotlib.use("Agg")
+import matplotlib.pyplot as plt
+import matplotlib.patches as mpatches
+import numpy as np
+import pandas as pd
+import trimesh
+
+
+# Camera convention: R,T are world-to-camera  →  center = -R^T @ T
+
+def load_cameras(calib: dict):
+    """Return arrays of camera positions, forward dirs, and IDs."""
+    positions, forwards, ids = [], [], []
+    for cam_id in sorted(calib.keys()):
+        data = calib[cam_id]
+        R = np.array(data["R"], dtype=np.float64).reshape(3, 3)
+        T = np.array(data["T"], dtype=np.float64).reshape(3)
+        c = -R.T @ T
+        fwd = R.T @ np.array([0.0, 0.0, 1.0])
+        positions.append(c)
+        forwards.append(fwd)
+        ids.append(cam_id)
+    return np.array(positions), np.array(forwards), ids
+
+
+def load_ego_trajectory(root, row):
+    """Load egocentric camera trajectory."""
+    ego_dir = row.get("ego_camera_params_dir", "")
+    if not ego_dir or pd.isna(ego_dir):
+        return None
+    ext_path = root / ego_dir / "egocentric_frame_extrinsic.npy"
+    if not ext_path.exists():
+        return None
+    exts = np.load(str(ext_path))
+    centers = []
+    for ext in exts:
+        R, T = ext[:3, :3], ext[:3, 3]
+        centers.append(-R.T @ T)
+    return np.array(centers)
+
+
+def load_objects(root, row, frame_idx=0):
+    """Load object mesh vertices in world coords at given frame."""
+    obj_dir = row.get("object_poses_dir", "")
+    if not obj_dir or pd.isna(obj_dir):
+        return {}
+    models_dir = root / "object_models_released"
+    obj_path = root / obj_dir
+    objects = {}
+    for npy_file in sorted(obj_path.glob("*.npy")):
+        poses = np.load(str(npy_file))
+        if frame_idx >= len(poses):
+            continue
+        pose = poses[frame_idx].astype(np.float64)
+        parts = npy_file.stem.split("_")
+        role, obj_id = parts[0], parts[1]
+        model_path = models_dir / f"{obj_id}_cm.obj"
+        if model_path.exists():
+            mesh = trimesh.load(str(model_path), process=False)
+            verts = np.array(mesh.vertices, dtype=np.float64) / 100.0  # cm→m
+            verts_world = (pose[:3, :3] @ verts.T).T + pose[:3, 3]
+            objects[f"{role}_{obj_id}"] = verts_world
+    return objects
+
+
+def plot_scene_2d(calib, root, row, save_path, frame_idx=0):
+    """Create a 2-panel (XY + XZ) plot."""
+    positions, forwards, cam_ids = load_cameras(calib)
+    ego_traj = load_ego_trajectory(root, row)
+    objects = load_objects(root, row, frame_idx)
+
+    fig, (ax_xy, ax_xz) = plt.subplots(1, 2, figsize=(18, 8))
+
+    cmap = plt.cm.tab20(np.linspace(0, 1, 12))
+    arrow_len = 0.15
+
+    # ── XY (top-down) ──
+    for i, (pos, fwd, cid) in enumerate(zip(positions, forwards, cam_ids)):
+        ax_xy.scatter(pos[0], pos[1], c=[cmap[i]], s=60, zorder=5, edgecolors="k", linewidth=0.5)
+        ax_xy.annotate(cid[-3:], (pos[0], pos[1]), fontsize=6, ha="center", va="bottom",
+                       xytext=(0, 5), textcoords="offset points", color=cmap[i])
+        ax_xy.arrow(pos[0], pos[1], fwd[0] * arrow_len, fwd[1] * arrow_len,
+                    head_width=0.03, head_length=0.02, fc=cmap[i], ec=cmap[i], alpha=0.7)
+
+    if ego_traj is not None:
+        ax_xy.plot(ego_traj[:, 0], ego_traj[:, 1], "m-", linewidth=1, alpha=0.5, label="ego")
+        ax_xy.scatter(ego_traj[frame_idx, 0], ego_traj[frame_idx, 1],
+                      c="magenta", marker="^", s=80, zorder=6, label=f"ego f{frame_idx}")
+
+    obj_colors = {"tool": "dodgerblue", "target": "coral"}
+    for name, verts in objects.items():
+        role = name.split("_")[0]
+        c = obj_colors.get(role, "gray")
+        # Plot bounding box outline
+        mn, mx = verts.min(axis=0), verts.max(axis=0)
+        rect = plt.Rectangle((mn[0], mn[1]), mx[0] - mn[0], mx[1] - mn[1],
+                              fill=True, facecolor=c, alpha=0.3, edgecolor=c, linewidth=1.5)
+        ax_xy.add_patch(rect)
+        centroid = verts.mean(axis=0)
+        ax_xy.text(centroid[0], centroid[1], name, fontsize=7, color=c, ha="center", va="center",
+                   fontweight="bold")
+
+    ax_xy.set_xlabel("X (m)")
+    ax_xy.set_ylabel("Y (m)")
+    ax_xy.set_title("Top-down (XY plane)")
+    ax_xy.set_aspect("equal")
+    ax_xy.grid(True, alpha=0.3)
+    ax_xy.legend(fontsize=7, loc="upper right")
+
+    # ── XZ (side view) ──
+    for i, (pos, fwd, cid) in enumerate(zip(positions, forwards, cam_ids)):
+        ax_xz.scatter(pos[0], pos[2], c=[cmap[i]], s=60, zorder=5, edgecolors="k", linewidth=0.5)
+        ax_xz.annotate(cid[-3:], (pos[0], pos[2]), fontsize=6, ha="center", va="bottom",
+                       xytext=(0, 5), textcoords="offset points", color=cmap[i])
+        ax_xz.arrow(pos[0], pos[2], fwd[0] * arrow_len, fwd[2] * arrow_len,
+                    head_width=0.03, head_length=0.02, fc=cmap[i], ec=cmap[i], alpha=0.7)
+
+    if ego_traj is not None:
+        ax_xz.plot(ego_traj[:, 0], ego_traj[:, 2], "m-", linewidth=1, alpha=0.5)
+        ax_xz.scatter(ego_traj[frame_idx, 0], ego_traj[frame_idx, 2],
+                      c="magenta", marker="^", s=80, zorder=6)
+
+    for name, verts in objects.items():
+        role = name.split("_")[0]
+        c = obj_colors.get(role, "gray")
+        mn, mx = verts.min(axis=0), verts.max(axis=0)
+        rect = plt.Rectangle((mn[0], mn[2]), mx[0] - mn[0], mx[2] - mn[2],
+                              fill=True, facecolor=c, alpha=0.3, edgecolor=c, linewidth=1.5)
+        ax_xz.add_patch(rect)
+        centroid = verts.mean(axis=0)
+        ax_xz.text(centroid[0], centroid[2], name, fontsize=7, color=c, ha="center", va="center",
+                   fontweight="bold")
+
+    ax_xz.set_xlabel("X (m)")
+    ax_xz.set_ylabel("Z (m)")
+    ax_xz.set_title("Side view (XZ plane)")
+    ax_xz.set_aspect("equal")
+    ax_xz.grid(True, alpha=0.3)
+
+    fig.suptitle(f"{row['sequence_id']}  (frame {frame_idx})", fontsize=13, fontweight="bold")
+    plt.tight_layout()
+    fig.savefig(save_path, dpi=150, bbox_inches="tight")
+    plt.close(fig)
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--csv", type=Path,
+                        default=Path(__file__).resolve().parent.parent / "taco_dataset" / "taco_info.csv")
+    parser.add_argument("--root", type=Path,
+                        default=Path(__file__).resolve().parent.parent / "taco_dataset")
+    parser.add_argument("--save-dir", type=Path,
+                        default=Path(__file__).resolve().parent / "plots" / "scene_2d")
+    parser.add_argument("--n-sequences", type=int, default=5)
+    parser.add_argument("--seed", type=int, default=42)
+    args = parser.parse_args()
+
+    df = pd.read_csv(args.csv)
+    df = df[df["all_modalities_complete"]].reset_index(drop=True)
+    args.save_dir.mkdir(parents=True, exist_ok=True)
+    root = args.root.resolve()
+
+    rng = np.random.RandomState(args.seed)
+    actions = sorted(df["action"].unique())
+    rng.shuffle(actions)
+
+    sampled = []
+    for action in actions:
+        subset = df[df["action"] == action]
+        idx = subset.index[rng.randint(0, len(subset))]
+        sampled.append(idx)
+        if len(sampled) >= args.n_sequences:
+            break
+
+    for i, seq_idx in enumerate(sampled):
+        row = df.iloc[seq_idx]
+        seq_id = row["sequence_id"]
+        print(f"[{i+1}/{len(sampled)}] {seq_id}")
+
+        calib_path = root / row["alloc_camera_params_path"]
+        with open(calib_path) as f:
+            calib = json.load(f)
+
+        safe = seq_id.replace("/", "_").replace(" ", "").replace(",", "").replace("(", "").replace(")", "")
+        plot_scene_2d(calib, root, row, args.save_dir / f"{safe}.png")
+        print(f"  saved {safe}.png")
+
+    print(f"\nDone — {len(sampled)} images saved to {args.save_dir}/")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/taco_analysis/visualize_taco_epipolar.py b/taco_analysis/visualize_taco_epipolar.py
new file mode 100644
index 0000000000000000000000000000000000000000..c507d3f139373831414ad1718ff3dbddf433e6d7
--- /dev/null
+++ b/taco_analysis/visualize_taco_epipolar.py
@@ -0,0 +1,335 @@
+#!/usr/bin/env python
+"""Visualize epipolar lines across TACO allocentric cameras.
+
+For a given sequence and frame, picks one source camera and draws random
+sample points on it.  For every other camera, computes the fundamental
+matrix from the allocentric calibration and draws the corresponding
+epipolar lines.  Arranges all views in a grid and saves the image.
+
+Usage:
+    python visualize_taco_epipolar.py \
+        --csv ../taco_dataset/taco_info.csv \
+        --root ../taco_dataset \
+        --save-dir plots/epipolar \
+        --n-sequences 3
+"""
+
+import argparse
+import json
+import math
+from pathlib import Path
+
+import imageio
+import numpy as np
+import pandas as pd
+from PIL import Image, ImageDraw, ImageFont
+
+
+# ── Epipolar geometry ────────────────────────────────────────────────
+
+def compute_fundamental_matrix(K_src, K_tgt, R_src, T_src, R_tgt, T_tgt):
+    """Compute F such that  x_tgt^T @ F @ x_src = 0.
+
+    Calibration gives world-to-camera transforms per camera:
+        p_cam = R @ p_world + T
+
+    So c2w_src = [R_src | T_src] inverted, etc.
+    Relative transform:  T_tgt_from_src = (R_tgt, T_tgt) @ inv(R_src, T_src)
+    """
+    # Relative rotation / translation
+    R_rel = R_tgt @ R_src.T                       # (3,3)
+    t_rel = T_tgt - R_rel @ T_src                 # (3,)
+
+    # Essential → Fundamental
+    tx = np.array([
+        [0, -t_rel[2], t_rel[1]],
+        [t_rel[2], 0, -t_rel[0]],
+        [-t_rel[1], t_rel[0], 0],
+    ], dtype=np.float64)
+
+    E = tx @ R_rel
+    K_tgt_inv_T = np.linalg.inv(K_tgt).T
+    K_src_inv = np.linalg.inv(K_src)
+    F = K_tgt_inv_T @ E @ K_src_inv
+    norm = np.linalg.norm(F)
+    return F / norm if norm > 1e-12 else F
+
+
+def epipolar_line_endpoints(F, px, py, W, H):
+    """Compute clipped endpoints of the epipolar line in the target image."""
+    line = F @ np.array([px, py, 1.0])
+    a, b, c = line
+
+    pts = []
+    if abs(b) > 1e-10:
+        y = -c / b
+        if -1 <= y <= H + 1:
+            pts.append((0.0, float(y)))
+        y = -(a * W + c) / b
+        if -1 <= y <= H + 1:
+            pts.append((float(W), float(y)))
+    if abs(a) > 1e-10:
+        x = -c / a
+        if -1 <= x <= W + 1:
+            pts.append((float(x), 0.0))
+        x = -(b * H + c) / a
+        if -1 <= x <= W + 1:
+            pts.append((float(x), float(H)))
+
+    unique = []
+    for p in pts:
+        if not any(abs(p[0] - u[0]) < 2 and abs(p[1] - u[1]) < 2 for u in unique):
+            unique.append(p)
+    return (unique[0], unique[1]) if len(unique) >= 2 else None
+
+
+# ── Colour helpers ───────────────────────────────────────────────────
+
+def random_colors(n, seed=42):
+    rng = np.random.RandomState(seed)
+    colors = []
+    for _ in range(n):
+        hue = rng.uniform(0, 1)
+        r = int(255 * max(0, min(1, abs(hue * 6 - 3) - 1)))
+        g = int(255 * max(0, min(1, 2 - abs(hue * 6 - 2))))
+        b = int(255 * max(0, min(1, 2 - abs(hue * 6 - 4))))
+        brightness = max(r, g, b)
+        if brightness < 150:
+            scale = 150 / max(brightness, 1)
+            r = min(255, int(r * scale))
+            g = min(255, int(g * scale))
+            b = min(255, int(b * scale))
+        colors.append((r, g, b))
+    return colors
+
+
+# ── Data loading ─────────────────────────────────────────────────────
+
+def load_calibration(calib_path: Path) -> dict:
+    """Load calibration.json → dict[cam_id] = {K, R, T}."""
+    with open(calib_path) as f:
+        raw = json.load(f)
+    calib = {}
+    for cam_id, data in raw.items():
+        K = np.array(data["K"], dtype=np.float64).reshape(3, 3)
+        R = np.array(data["R"], dtype=np.float64).reshape(3, 3)
+        T = np.array(data["T"], dtype=np.float64).reshape(3,)
+        calib[cam_id] = {"K": K, "R": R, "T": T}
+    return calib
+
+
+def read_frame(video_path: Path, frame_idx: int) -> np.ndarray:
+    """Read a single frame from an MP4 video via imageio."""
+    reader = imageio.get_reader(str(video_path))
+    frame = reader.get_data(frame_idx)
+    reader.close()
+    return frame  # (H, W, 3) uint8
+
+
+# ── Grid visualisation ───────────────────────────────────────────────
+
+def build_epipolar_grid(
+    root: Path,
+    row: pd.Series,
+    frame_idx: int,
+    source_cam_idx: int = 0,
+    num_points: int = 15,
+    seed: int = 42,
+    thumb_width: int = 640,
+) -> Image.Image:
+    """Build a grid image for one sequence / frame.
+
+    The source camera gets random sample points drawn.
+    Every other camera gets the corresponding epipolar lines.
+    """
+    camera_ids = row["camera_ids"].split(";")
+    n_cams = len(camera_ids)
+    source_cam = camera_ids[source_cam_idx]
+
+    # Load calibration
+    calib_path = root / row["alloc_camera_params_path"]
+    calib = load_calibration(calib_path)
+
+    # Load source frame
+    mr_dir = root / row["marker_removed_dir"]
+    src_frame = read_frame(mr_dir / f"{source_cam}.mp4", frame_idx)
+    H_orig, W_orig = src_frame.shape[:2]
+
+    # Thumbnail dimensions (preserve aspect)
+    thumb_h = int(thumb_width * H_orig / W_orig)
+    scale_x = thumb_width / W_orig
+    scale_y = thumb_h / H_orig
+
+    # Sample random points on the source image
+    rng = np.random.RandomState(seed)
+    margin_x, margin_y = W_orig // 6, H_orig // 6
+    pts = [(rng.randint(margin_x, W_orig - margin_x),
+            rng.randint(margin_y, H_orig - margin_y))
+           for _ in range(num_points)]
+    colors = random_colors(num_points, seed)
+
+    K_src = calib[source_cam]["K"]
+    R_src = calib[source_cam]["R"]
+    T_src = calib[source_cam]["T"]
+
+    # Build each camera tile
+    tiles = []
+    label_h = 24
+    for cam_id in camera_ids:
+        frame = read_frame(mr_dir / f"{cam_id}.mp4", frame_idx)
+        img = Image.fromarray(frame).resize((thumb_width, thumb_h), Image.LANCZOS)
+        draw = ImageDraw.Draw(img)
+
+        point_r = max(3, thumb_width // 120)
+        line_w = max(2, thumb_width // 240)
+
+        if cam_id == source_cam:
+            # Draw sample points on source
+            for (px, py), color in zip(pts, colors):
+                sx, sy = px * scale_x, py * scale_y
+                draw.ellipse(
+                    [sx - point_r, sy - point_r, sx + point_r, sy + point_r],
+                    fill=color, outline=(255, 255, 255), width=1,
+                )
+        else:
+            # Draw epipolar lines on target
+            K_tgt = calib[cam_id]["K"]
+            R_tgt = calib[cam_id]["R"]
+            T_tgt = calib[cam_id]["T"]
+            F = compute_fundamental_matrix(K_src, K_tgt, R_src, T_src, R_tgt, T_tgt)
+            for (px, py), color in zip(pts, colors):
+                endpoints = epipolar_line_endpoints(F, px, py, W_orig, H_orig)
+                if endpoints:
+                    (x1, y1), (x2, y2) = endpoints
+                    draw.line(
+                        [(x1 * scale_x, y1 * scale_y),
+                         (x2 * scale_x, y2 * scale_y)],
+                        fill=color, width=line_w,
+                    )
+                # Small dot at same pixel location (for reference)
+                sx, sy = px * scale_x, py * scale_y
+                r2 = max(2, point_r // 2)
+                draw.ellipse(
+                    [sx - r2, sy - r2, sx + r2, sy + r2],
+                    fill=None, outline=color, width=1,
+                )
+
+        # Add label bar on top
+        tile = Image.new("RGB", (thumb_width, thumb_h + label_h), (30, 30, 30))
+        tile_draw = ImageDraw.Draw(tile)
+        label = f"cam {cam_id}"
+        if cam_id == source_cam:
+            label += "  [SOURCE]"
+            tile_draw.text((4, 4), label, fill=(100, 255, 100))
+        else:
+            tile_draw.text((4, 4), label, fill=(200, 200, 200))
+        tile.paste(img, (0, label_h))
+        tiles.append(tile)
+
+    # Arrange in grid: 4 columns × 3 rows (for 12 cameras)
+    n_cols = 4
+    n_rows = math.ceil(n_cams / n_cols)
+    tile_w, tile_h = tiles[0].size
+    gap = 3
+
+    # Title bar
+    title_h = 36
+    grid_w = n_cols * tile_w + (n_cols - 1) * gap
+    grid_h = title_h + n_rows * tile_h + (n_rows - 1) * gap
+    grid = Image.new("RGB", (grid_w, grid_h), (20, 20, 20))
+    grid_draw = ImageDraw.Draw(grid)
+
+    title = f"{row['sequence_id']}  |  frame {frame_idx}  |  source: cam {source_cam}  |  {num_points} points"
+    grid_draw.text((10, 8), title, fill=(255, 255, 255))
+
+    for i, tile in enumerate(tiles):
+        r, c = divmod(i, n_cols)
+        x = c * (tile_w + gap)
+        y = title_h + r * (tile_h + gap)
+        grid.paste(tile, (x, y))
+
+    return grid
+
+
+# ── Main ─────────────────────────────────────────────────────────────
+
+def main():
+    parser = argparse.ArgumentParser(description="Visualize TACO epipolar lines across cameras")
+    parser.add_argument("--csv", type=Path,
+                        default=Path(__file__).resolve().parent.parent / "taco_dataset" / "taco_info.csv")
+    parser.add_argument("--root", type=Path,
+                        default=Path(__file__).resolve().parent.parent / "taco_dataset")
+    parser.add_argument("--save-dir", type=Path,
+                        default=Path(__file__).resolve().parent / "plots" / "epipolar")
+    parser.add_argument("--n-sequences", type=int, default=5,
+                        help="Number of sequences to sample")
+    parser.add_argument("--frames-per-seq", type=int, default=2,
+                        help="Frames to visualize per sequence")
+    parser.add_argument("--num-points", type=int, default=15)
+    parser.add_argument("--seed", type=int, default=42)
+    parser.add_argument("--thumb-width", type=int, default=512)
+    args = parser.parse_args()
+
+    df = pd.read_csv(args.csv)
+    # Filter to complete sequences with all cameras
+    df = df[(df["all_modalities_complete"]) & (df["n_allocentric_cameras"] == 12)].reset_index(drop=True)
+    print(f"Loaded {len(df)} complete sequences with 12 cameras")
+
+    args.save_dir.mkdir(parents=True, exist_ok=True)
+
+    # Sample diverse sequences (one per action, round-robin)
+    rng = np.random.RandomState(args.seed)
+    actions = sorted(df["action"].unique())
+    rng.shuffle(actions)
+    sampled_indices = []
+    action_cycle = 0
+    while len(sampled_indices) < args.n_sequences:
+        action = actions[action_cycle % len(actions)]
+        subset = df[df["action"] == action]
+        pick = rng.randint(0, len(subset))
+        global_idx = subset.index[pick]
+        if global_idx not in sampled_indices:
+            sampled_indices.append(global_idx)
+        action_cycle += 1
+    sampled_indices = sampled_indices[:args.n_sequences]
+
+    # Also cycle through different source cameras
+    camera_ids_all = df.iloc[0]["camera_ids"].split(";")
+
+    total = 0
+    for i, seq_idx in enumerate(sampled_indices):
+        row = df.iloc[seq_idx]
+        n_frames = int(row["n_frames"])
+        seq_id_safe = row["sequence_id"].replace("/", "_").replace(" ", "").replace(",", "").replace("(", "").replace(")", "")
+
+        # Pick evenly-spaced frames
+        if args.frames_per_seq >= n_frames:
+            frame_indices = list(range(n_frames))
+        else:
+            frame_indices = np.linspace(0, n_frames - 1, args.frames_per_seq, dtype=int).tolist()
+
+        # Rotate source camera across sequences
+        source_cam_idx = i % len(camera_ids_all)
+
+        print(f"\n[{i+1}/{args.n_sequences}] {row['sequence_id']}  ({n_frames} frames, source cam idx={source_cam_idx})")
+        for fidx in frame_indices:
+            print(f"  frame {fidx}...", end=" ", flush=True)
+            grid = build_epipolar_grid(
+                root=args.root.resolve(),
+                row=row,
+                frame_idx=fidx,
+                source_cam_idx=source_cam_idx,
+                num_points=args.num_points,
+                seed=args.seed + fidx,
+                thumb_width=args.thumb_width,
+            )
+            fname = f"{seq_id_safe}_f{fidx:04d}_src{source_cam_idx}.png"
+            grid.save(args.save_dir / fname)
+            print(f"saved {fname}")
+            total += 1
+
+    print(f"\nDone — {total} grid images saved to {args.save_dir}/")
+
+
+if __name__ == "__main__":
+    main()