# Table extraction from scrolling video

Extract a table from a screen-recorded video: sample frames, OCR (Portuguese), align to column/row bounds from an SVG template, then merge and deduplicate into one CSV.

## Inputs

| File | Role |
|------|------|
| `video-data.webm` | Source video (scroll-down table). |
| `template.svg` | Annotation: rectangles under the image = column bounds; 28 rows. |

## Outputs

| File / dir | Role |
|------------|------|
| `result.csv` | Final table: 6 columns (Concessionária, Código, Rodovia/UF, km inicial, km final, Extensão), one row per segment, deduplicated. |
| `frames/` | Working: one PNG and one CSV per sampled frame (e.g. 0, 10, 20, …). |
| `llm-fixed-frames/` | Optional: copy of frame CSVs after manual/LLM fixes; sew from here instead of `frames/` if used. |

## Run (same chore again)

**Fully automated (extract → fix → sew → result):**

```bash
nix-shell -p python3 python3Packages.opencv4 python3Packages.numpy python3Packages.pytesseract tesseract ffmpeg --run "./run.sh"
```

Or with custom video / output dir:

```bash
./run.sh path/to/video.webm my_frames
# result.csv is still written at project root
```

**If you do manual fixes:** copy `frames/*.csv` into `llm-fixed-frames/`, edit CSVs, then:

```bash
python3 sew_csvs.py llm-fixed-frames result.csv
```

## Scripts (keep)

| Script | Role |
|--------|------|
| `extract_frames_and_tables.py` | Sample video every N frames → PNGs; OCR (por) + SVG column/row bounds → one CSV per frame. |
| `fix_all_csvs.py` | Heuristic fixes on frame CSVs (strip, E→-, pipe→space, extensão from km). |
| `sew_csvs.py` | Merge frame CSVs in order, remove boundary overlap, deduplicate full rows, write result. |
| `svg_columns.py` | Parse column/row rectangles from template.svg. |
| `assign_cells.py` | Map word boxes to (col, row); merge cell text (col1 no space, others space). |
| `clean_csv_heuristics.py` | Per-row cleanup and extensão correction (used by fix_all_csvs). |
| `row_eq.py` | Row equality for sewing. |

## Not kept (removed)

- `extract_table_frames.py` – superseded by `extract_frames_and_tables.py`.
- `extract_every_n_frames.py` – logic folded into `extract_frames_and_tables.py`.
- `sewn.csv` – superseded by `result.csv` (sew now writes result.csv and dedupes).

## Clean re-run

To start from scratch:

- Delete or clear `frames/` and optionally `llm-fixed-frames/`.
- Run `./run.sh` (or the manual-fix flow above).

`result.csv` is overwritten each run.