---
license: mit
task_categories:
  - object-detection
  - video-classification
language:
  - en
tags:
  - car-parts
  - semantic-retrieval
  - yolo
  - automotive
pretty_name: RAV4 Video Detection and Retrieval Index
configs:
  - config_name: detections
    data_files:
      - split: detections
        path: video_detections.parquet
  - config_name: retrieval
    data_files:
      - split: retrieval
        path: retrieval_output.parquet
---

# RAV4 Video Detection and Retrieval Index

This dataset contains two Parquet files produced by a YOLO26 object detection pipeline
applied to a Toyota RAV4 exterior review video. Together they form a complete
image-to-video semantic retrieval system.

## Source Video

| Field         | Value                                               |
|---------------|-----------------------------------------------------|
| Video ID      | `YcvECxtXoxQ`                                       |
| YouTube URL   | https://www.youtube.com/watch?v=YcvECxtXoxQ         |
| Sampling rate | 0.5 fps (one frame every 2 seconds)                 |
| Total frames  | ~1,396                                              |

## Detector

| Field            | Value                                                       |
|------------------|-------------------------------------------------------------|
| Model            | YOLO26n-seg (nano segmentation variant)                     |
| Base weights     | `yolo26n-seg.pt` (COCO pretrained)                          |
| Fine-tuning data | Ultralytics Car Parts Segmentation Dataset (`carparts-seg.yaml`) |
| Training epochs  | 50                                                          |
| Image size       | 640 × 640                                                   |
| Batch size       | 32                                                          |
| Hardware         | NVIDIA A100 80GB (Google Colab)                             |
| Confidence cutoff| 0.30                                                        |

## Files

| File                       | Description                                        |
|----------------------------|----------------------------------------------------|
| `video_detections.parquet` | Detection index — one row per video frame          |
| `retrieval_output.parquet` | Retrieval results — one row per query × class pair |

---

## `video_detections.parquet` Schema

One row per video frame. All bounding-box detections within a frame are grouped
as parallel lists, so `class_labels[i]`, `confidence_scores[i]`, `x_mins[i]`, etc.
all correspond to the same detection.

| Column              | Type          | Description                                         |
|---------------------|---------------|-----------------------------------------------------|
| `video_id`          | string        | YouTube video ID (`YcvECxtXoxQ`)                    |
| `frame_index`       | int32         | 1-based frame number from ffmpeg output             |
| `timestamp_sec`     | float32       | Wall-clock position: `frame_index × 2` seconds      |
| `detection_count`   | int32         | Number of detections in this frame (0 if none)      |
| `class_labels`      | list[string]  | Detected car part names (see class list below)      |
| `confidence_scores` | list[float32] | YOLO confidence per detection, in [0, 1]            |
| `x_mins`            | list[float32] | Bounding box left edges in pixels                   |
| `y_mins`            | list[float32] | Bounding box top edges in pixels                    |
| `x_maxs`            | list[float32] | Bounding box right edges in pixels                  |
| `y_maxs`            | list[float32] | Bounding box bottom edges in pixels                 |
| `detector_name`     | string        | Model identifier: `"yolo26n-seg-carparts-ft50ep"`   |

---

## `retrieval_output.parquet` Schema

One row per (query image × detected class label) pair. All retrieved video segments
for that combination are grouped as parallel lists, so `start_timestamps[i]`,
`end_timestamps[i]`, `supporting_detections[i]`, and `verification_urls[i]` all
describe the same segment.

| Column                       | Type          | Description                                              |
|------------------------------|---------------|----------------------------------------------------------|
| `query_timestamp_sec`        | int32         | Source timestamp of the query image in seconds           |
| `query_timestamp`            | string        | Human-readable "MM:SS" label                             |
| `query_class_labels`         | string        | Comma-separated list of all classes detected in query    |
| `class_label`                | string        | The class driving this retrieval group                   |
| `segment_count`              | int32         | Number of contiguous segments found in the video         |
| `start_timestamps`           | list[float32] | Segment start times in the corpus video (seconds)        |
| `end_timestamps`             | list[float32] | Segment end times in the corpus video (seconds)          |
| `supporting_detections`      | list[int32]   | Number of supporting frame detections per segment        |
| `total_supporting_detections`| int32         | Sum of supporting_detections across all segments         |
| `verification_urls`          | list[string]  | YouTube embed URLs for visual verification per segment   |

---

## Class Labels

The 23 classes come from `carparts-seg.yaml`:

| ID | Label               | ID | Label               |
|----|---------------------|----|---------------------|
| 0  | `back_bumper`       | 12 | `front_left_light`  |
| 1  | `back_door`         | 13 | `front_light`       |
| 2  | `back_glass`        | 14 | `front_right_door`  |
| 3  | `back_left_door`    | 15 | `front_right_light` |
| 4  | `back_left_light`   | 16 | `hood`              |
| 5  | `back_light`        | 17 | `left_mirror`       |
| 6  | `back_right_door`   | 18 | `object`            |
| 7  | `back_right_light`  | 19 | `right_mirror`      |
| 8  | `front_bumper`      | 20 | `tailgate`          |
| 9  | `front_door`        | 21 | `trunk`             |
| 10 | `front_glass`       | 22 | `wheel`             |
| 11 | `front_left_door`   |    |                     |

---

## How to Load

```python
import pandas as pd

# Detection index
detections = pd.read_parquet(
    "hf://datasets/ccczy-czy/rav4-video-detections/video_detections.parquet"
)

# Retrieval results
retrieval = pd.read_parquet(
    "hf://datasets/ccczy-czy/rav4-video-detections/retrieval_output.parquet"
)
```

## Example: Find All Segments Where a Class Was Detected

```python
import pandas as pd

retrieval = pd.read_parquet(
    "hf://datasets/ccczy-czy/rav4-video-detections/retrieval_output.parquet"
)

# All retrieval groups for front_left_light
rows = retrieval[retrieval["class_label"] == "front_left_light"]
for _, row in rows.iterrows():
    for start, end, url in zip(
        row["start_timestamps"], row["end_timestamps"], row["verification_urls"]
    ):
        print(f"  {start:.0f}s – {end:.0f}s  {url}")
```