# Paper Figures — Review Page

Status: figures being generated by visualizations agent. Refresh to see updates.

---

<details open>
<summary><h2>Figure 1: Overview</h2></summary>

**Goal:** Three-panel teaser — core idea + three domains + headline numbers.

**Status:** Pending

![Fig 1](/paper_figures/media/fig1_overview.png)

</details>

---

<details open>
<summary><h2>Figure 2a: Pixel-Aligned Heatmap Volume pipeline</h2></summary>

**Goal:** RGB → DINO → features → 1×1 conv → 32 ray logits → unproject → 3 viewers.

![Fig 2a](/paper_figures/media/fig2a_method.png)

[Download fig2a_method.svg](media/fig2a_method.svg)

</details>

---

<details open>
<summary><h2>Figure 2b: Height prediction is view-invariant</h2></summary>

**Goal:** Two camera frustums on the same 3D target, different depths, same height z*.

![Fig 2b](/paper_figures/media/fig2b_invariance.png)

[Download fig2b_invariance.svg](media/fig2b_invariance.svg)

</details>

---

<details open>
<summary><h2>Figure 3: Real Robot Results</h2></summary>

**Goal:** 3 tasks + 3 OOD conditions with photos and success rates.

**Status:** Pending

![Fig 3](/paper_figures/media/fig3_realrobot.png)

</details>

---

<details open>
<summary><h2>Figure 4: OOD Generalization Analysis</h2></summary>

**Goal:** 2×3 layout — (a) spatial generalization (distribution / chart / qualitative) and (b) viewpoint generalization (polar / per-theta chart / qualitative).

![Fig 4](/paper_figures/media/fig4_ood.png)

[Download fig4_ood.svg](media/fig4_ood.svg)

</details>

---

<details open>
<summary><h2>Figure 5: Video Backbone</h2></summary>

**Goal:** Two-stage training diagram + rollout grid comparison (92% vs 0%).

**Status:** Pending

![Fig 5](/paper_figures/media/fig5_video.png)

</details>

---

<details open>
<summary><h2>Figure 6: Point Track Pretraining</h2></summary>

**Goal:** Arm-deleted training data + results bar chart.

**Status:** Pending

![Fig 6](/paper_figures/media/fig6_pretrain.png)

</details>

---

## SVG versions (Figma-editable)

Pure hand-authored SVG versions of the diagram-heavy figures. Every element is
a real `<rect>` / `<text>` / `<line>` / `<image>` — import into Figma (File → Place,
or drag-drop onto the canvas) and you get an editable vector group with Inter
text that Figma renders natively. Photos/video frames are embedded as base64
so the SVG files are fully self-contained.

### Fig 1: Overview (SVG)

![Fig 1 SVG preview](media/fig1_overview_svg.png)

[Download fig1_overview.svg](media/fig1_overview.svg)

### Fig 2: Method (SVG)

![Fig 2 SVG preview](media/fig2_method_svg.png)

[Download fig2_method.svg](media/fig2_method.svg)

### Fig 5: Video Backbone (SVG)

![Fig 5 SVG preview](media/fig5_video_svg.png)

[Download fig5_video.svg](media/fig5_video.svg)

HTML/CSS source files remain at `/data/cameron/para/paper/figs/html/` as a
fallback — the HTML-rendered PNGs (without `_svg` suffix) are unchanged.

---

## Checklist

- [x] **Fig 1**: Overview teaser — **shipped (HTML + CSS via Playwright)**, Inter typography, three-panel polished layout
- [x] **Fig 2**: Method pipeline — shipped (matplotlib composite of cleaned video frames + arrows + height illustration); HTML version also prepared at `paper/figs/html/fig2_method.html` (not rendered, optional upgrade)
- [x] **Fig 3**: Real robot results — shipped (single image import from Figma)
- [x] **Fig 4**: OOD Analysis composite — shipped (stacked layout: 4a + 4b on left, 4c on right)
  - [x] **Fig 4a**: Spatial extrapolation (distribution plot + 54%/1% bar chart)
  - [x] **Fig 4b**: Per-theta viewpoint chart — matplotlib PNG + SVG at dpi 300
  - [x] **Fig 4c**: Qualitative frames with red/green annotations on ACT/PARA mid-rollout
- [x] **Fig 5**: Video backbone — **shipped (HTML + CSS via Playwright)**, clean two-stage pipeline, color-coded rollout cards
- [ ] **Fig 6**: Point track pretraining — placeholder gray rects in place; arm-deleted source frames already extracted to `/data/cameron/penpot/figures/extracted/fig6_arm_deleted_{0..4}.png`, ready to swap once backbones agent finishes the experiment

### Infrastructure
- [x] All figures rendered at dpi 300 (matplotlib) or 2× device pixel ratio (Playwright) for print quality
- [x] All figures saved to both `paper/figs/generated/` (LaTeX) and `.agents/reports/paper_figures/media/` (web review)
- [x] Review page live at https://omidlab.net/paper_figures/
- [x] HTML source files saved to `paper/figs/html/` for future iteration (Fig 1, Fig 2, Fig 5)
- [x] Render scripts:
  - `render_paper_figures.py` — matplotlib (Fig 2, 3, 4, 4a, 4b, 4c, 6)
  - `render_html_figures.py` — Playwright HTML→PNG (Fig 1, 2, 5)

**Status**: Core figure set complete. Fig 6 awaits real arm-deletion data from the backbones agent.