data/tracking/README.md

# prepare_sot_dataset.py

Prepares an evaluator-ready **Single Object Tracking (SOT)** dataset from a
benchmark JSON/JSONL file and source camera videos hosted on Hugging Face.

---

## Files

| File | Description |
|---|---|
| `sot_benchmark.jsonl` | Benchmark file defining all evaluation sequences — scene, camera, object, init bounding box, and canonical 8-frame IDs. |
| `prepare_sot_dataset.py` | Script that downloads source videos and extracts frames into an evaluator-ready directory. |

---

## Overview

The script:
1. Reads `sot_benchmark.jsonl` (or a custom benchmark file) describing SOT sequences (scene, camera, target bounding box, frame IDs).
2. Downloads the corresponding source `.mp4` files from the `nvidia/PhysicalAI-SmartSpaces` Hugging Face dataset (or a custom repo).
3. Extracts the requested frames with `ffmpeg`.
4. Annotates the initialization frame with a bounding-box overlay (`f00_ann.png`) and saves a cropped target thumbnail (`crop.png`).
5. Writes a `sequence_meta.json` per sequence.
6. Produces a `gt_requests.json` manifest. **Ground-truth bounding boxes are never written to the output directory** — see [GT Request Submission](#gt-request-submission) below.

---

## Prerequisites

| Requirement | Notes |
|---|---|
| Python 3.8+ | |
| `ffmpeg` | Must be on `PATH` (or in a few well-known locations). |
| `huggingface_hub` | `pip install huggingface_hub` |
| `opencv-python` **or** `Pillow` | Used for bbox drawing and cropping. `opencv-python` is preferred. |
| HF access token | Read access to the source video dataset. |

```bash

pip install huggingface_hub opencv-python

```

---

## Hugging Face Authentication

A token with **read** access to `nvidia/PhysicalAI-SmartSpaces` is required.
Pass it in one of two ways:

```bash

# Option A – command-line flag

--hf-token hf_xxxxxxxxxxxx



# Option B – environment variable (recommended for scripts/CI)

export HF_TOKEN=hf_xxxxxxxxxxxx

```

---

## Input Formats

The script accepts four flavors of JSON / JSONL input.

### 1. Standard benchmark JSON/JSONL *(most common)*

```json

[

  {

    "seq_id": "Warehouse_016__Camera_11__5600__obj353",

    "scene": "Warehouse_016",

    "camera": "Camera_11",

    "object_id": "353",

    "object_type": "Robot",

    "init_frame_id": 5600,

    "init_bbox": [799.0, 601.9, 918.8, 956.5],

    "canonical_frame_ids": [5600, 5615, 5630, 5645, 5660, 5675, 5690, 5705],

    "clip_fps": 30.0

  }

]

```

- `init_bbox` — normalized coordinates in **thousandths** of image dimensions
  (`[x1, y1, x2, y2]` where 1000 = full width/height).
- `canonical_frame_ids` — preferred source-video frame indices to extract.
  When provided and long enough, they take priority over the stride calculation.

### 2. Benchmark JSONL with explicit GT

```json

{"seq_id": "...", "canonical_frame_ids": [...], "gt_bboxes": {"5600": [...]}}

```

### 3. Dataset JSONL (metadata + conversations)

```json

{

  "id": "...",

  "metadata": {"scene": "...", "camera": "...", "init_bbox": [...], ...},

  "conversations": [{"role": "user", "value": "..."}, {"role": "assistant", "value": "{\"5600\": [...]}"}]

}

```

### 4. Sequence-only JSON/JSONL

```json

{"id": "...", "scene": "...", "camera": "...", "source_frame_ids": [...], "init_bbox": [...]}

```

---

## Output Structure

```

<output_dir>/

  gt_requests.json            # submit this to us to receive GT annotations (see below)

  <seq_id>/

    frames/

      f00.png                 # initialization frame

      f00_ann.png             # initialization frame with target bbox drawn

      crop.png                # cropped target region

      f01.png

      f02.png

      ...

      f{N-1}.png

    sequence_meta.json        # per-sequence metadata

```

### `sequence_meta.json` fields



| Field | Description |

|---|---|

| `frame_ids` | Source-video frame indices that were extracted |
| `init_bbox` | Target bounding box (thousandths) |
| `label` | Human-readable sequence label |
| `scene` / `camera` | Source identifiers |
| `object_id` / `object_type` | Target object metadata |
| `stride` | Frame stride used during extraction |
| `nframes` | Number of frames extracted |
| `clip_fps` | Frame rate of the source video |
| `gt_available` | Always `false` (GT is private) |

---

## Usage

### Minimal

```bash

python scripts/tracking/prepare_sot_dataset.py \

    --benchmark sot_benchmark.jsonl \

    --output-dir ./SOT_prepared_8f

```

### Extract 32 frames per sequence

```bash

python scripts/tracking/prepare_sot_dataset.py \

    --benchmark sot_benchmark.jsonl \

    --output-dir ./SOT_prepared_32f \

    --nframes 32

```

### Custom frame stride

```bash

python scripts/tracking/prepare_sot_dataset.py \

    --benchmark sot_benchmark.jsonl \

    --output-dir ./SOT_prepared \

    --nframes 16 \

    --stride 10

```

### Process only specific sequences

```bash

python scripts/tracking/prepare_sot_dataset.py \

    --benchmark sot_benchmark.jsonl \

    --output-dir ./SOT_prepared \

    --sequences Warehouse_016__Camera_11__5600__obj353 Warehouse_016__Camera_05__704__obj352

```

### Use a custom Hugging Face cache directory

```bash

python scripts/tracking/prepare_sot_dataset.py \

    --benchmark sot_benchmark.jsonl \

    --output-dir ./SOT_prepared \

    --hf-token hf_xxxxxxxxxxxx \

    --hf-cache-dir /data/hf_cache

```

### Windows (PowerShell)

```powershell

python scripts/tracking/prepare_sot_dataset.py `

    --benchmark sot_benchmark.jsonl `

    --output-dir .\SOT_prepared_8f `

    --hf-token hf_xxxxxxxxxxxx

```

---

## Command-Line Reference

| Argument | Required | Default | Description |
|---|---|---|---|
| `--benchmark` | Yes | — | Path to the input benchmark JSON or JSONL file. |
| `--output-dir` | Yes | — | Directory where prepared sequences are written. Created if it does not exist. |
| `--nframes` | No | `8` | Number of frames to extract per sequence. |
| `--stride` | No | auto | Source-video frame stride. When omitted, auto-computed from `clip_end_frame` or taken directly from `canonical_frame_ids`. |
| `--hf-token` | No* | `$HF_TOKEN` | Hugging Face access token. Falls back to the `HF_TOKEN` environment variable. |
| `--hf-cache-dir` | No | HF default | Cache directory for downloaded videos. |
| `--repo-id` | No | `nvidia/PhysicalAI-SmartSpaces` | Hugging Face dataset repository. |
| `--repo-subdir` | No | `MTMC_Tracking_2025` | Subdirectory inside the repository. |
| `--sequences` | No | all | Space-separated list of `seq_id` values to prepare. All others are skipped. |

\* Required in practice unless `HF_TOKEN` is set in the environment.

---

## Resuming an Interrupted Run

The script checks how many frames already exist in each sequence's `frames/`
directory. If the count equals or exceeds `--nframes`, that sequence is skipped
automatically. You can safely re-run the command after an interruption; only
incomplete sequences will be processed.

---

## Frame Stride Logic

Frames are selected using this priority order:

1. **`canonical_frame_ids`** in the benchmark — used directly when the list
   has at least `--nframes` entries and `--stride` is not explicitly set.
2. **`--stride`** — fixed stride supplied by the user.
3. **Auto** — computed as `min(15, (clip_end_frame - init_frame_id) / (nframes - 1))`,
   capped at the default stride of 15.

---

## Train / Val Split

The script infers the dataset split from the scene name:

- `Warehouse_000` – `Warehouse_014` → **train**
- `Warehouse_015` and above → **val**

This determines which subdirectory (`train/` or `val/`) is used when
constructing the download path on Hugging Face.

---

## GT Request Submission

The `sot_benchmark.jsonl` file defines the canonical **8-frame** evaluation sequences with fixed frame IDs. If you use the default settings (`--nframes 8` without `--stride`), no submission is needed — the benchmark frames are already known.

If you run a **custom variant** (e.g. `--nframes 16`, `--nframes 32`, `--nframes 64`, or a custom `--stride`), the script will produce a `gt_requests.json` file in your output directory once preparation is complete. **Submit this file back to us** so we can look up and return the ground-truth annotations for your chosen frames.

Send `gt_requests.json` to: *(benchmark contact TBD)*

---

## Evaluator Integration

After preparation, point your evaluator config at the output directory:

```json

{

  "prepared_data_dir": "./SOT_prepared_8f"

}

```

---

## Troubleshooting

| Problem | Fix |
|---|---|
| `ERROR: ffmpeg not found` | Install ffmpeg and add it to `PATH`, or place it in `/usr/bin/ffmpeg`. |
| `401 Unauthorized` from Hugging Face | Check that `--hf-token` / `HF_TOKEN` is set and has read access to the repo. |
| Download retries / timeouts | The script retries up to 4 times with back-off. Check your network connection. |
| Bounding boxes not drawn | Install `opencv-python` or `Pillow`. Without either, the script copies the raw frame without annotation. |
| Sequence not found in output | Verify the `seq_id` value; use `--sequences <seq_id>` to test a single entry. |