File size: 5,627 Bytes
2d5be72 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 | ---
pretty_name: OpenVO Trajectory Mid-Training Data
language:
- en
tags:
- autonomous-driving
- trajectory-prediction
- multimodal
configs:
- config_name: openvo_119w
data_files:
- split: train
path: annotations/openvo_navsim_vqa_all_processed_10deg_clean.jsonl
- config_name: openvo_54w
data_files:
- split: train
path: annotations/openvo_navsim_vqa_54w_10deg_clean.jsonl
- config_name: openvo_149w
data_files:
- split: train
path: annotations/openvo_navsim_vqa_149w_10deg_clean.jsonl
---
# OpenVO Trajectory Mid-Training Data
This dataset contains three original, geometry-cleaned OpenVO trajectory JSONL files and the JPEG images they reference. It supports the earlier 1,193,991-example experiment and the newer 2,031,368-example experiment. Images shared by the two experiments are stored once.
| Annotation set | Examples | Image directory after extraction |
|---|---:|---|
| Original 119w | 1,193,991 | `images/ScenePilot-dataset` |
| New 54w | 539,452 | `images/scene_pilot_2_dataset` |
| New 149w | 1,491,916 | `images/ScenePilot-dataset` |
There are **2,031,368 unique JPEG images**. The new 149w image set includes every image referenced by the original 119w annotations. The annotation files are retained independently: 4,949 overlapping records have different conversation text, including 4,774 with different answer text. These counts compare text without record IDs; they do not imply a difference in every numerical trajectory value.
For the original experiment, use the 119w JSONL only. For the new experiment, combine 54w and 149w, totaling 2,031,368 examples. Each file has independently numbered IDs starting at zero. The files are complete source annotation sets; the `train` labels in the dataset configurations do not define a benchmark train/validation split. Split by video directory when creating validation sets to avoid adjacent-frame leakage.
## Contents and data format
- `annotations/`: the three JSONL files, preserved byte for byte.
- `images/54w/` and `images/149w/`: uncompressed tar shards, each below 5 GB. Each member retains its JPEG bytes and portable relative path beneath `images/<image-directory>/`.
- `manifests/`: annotation contracts, cleaning/conversion provenance, per-image compressed JSONL indices (path, size, SHA-256), shard checksums, and exact package counts and sizes.
- `SHA256SUMS`: SHA-256 checksums of all distributed files except this checksum file itself.
- `source_roots.yaml`: portable annotation-to-image-root mapping and both experiment recipes.
- `verify_and_extract.py`: standard-library verification and extraction tool.
Each annotation is an object with `id`, `image`, and `conversations`. The `image` value is relative to the corresponding image root. Conversation roles are `system`, `human`, and `gpt`; the human prompt contains navigation/history context and the response contains the trajectory supervision. No training text, coordinates, navigation labels, images, or IDs were rewritten for this release. The dataset viewer configurations expose the JSONL records; download and extract the tar shards to resolve the relative image paths.
## Download and use on another server
Download the complete dataset using a Hugging Face client, for example:
```bash
hf download data-loader/OpenVO-Traj-midtrain --repo-type dataset --local-dir OpenVO-Traj-midtrain
python OpenVO-Traj-midtrain/verify_and_extract.py --destination OpenVO-Traj-unpacked
```
To check integrity without extracting:
```bash
python OpenVO-Traj-midtrain/verify_and_extract.py --verify-only
```
The extraction tool verifies the bundle before unpacking and requires a destination without existing `images` or `annotations` directories. Retaining both downloaded archives and extracted images requires roughly twice the package size; use `manifests/package_summary.json` for exact byte counts.
Set each loader's image root to the corresponding directory inside `OpenVO-Traj-unpacked`, and its annotation path to the matching JSONL inside `OpenVO-Traj-unpacked/annotations`. `source_roots.yaml` lists the relative mappings.
## Trajectory convention and cleaning
The reference point is the current front-camera center. Coordinates are camera-local, with x pointing forward and y left; positions are in meters and headings in radians. Records use four history points and six future points at 0.5-second intervals. The 10-degree navigation-command version is used for training; the source 20-degree version was used for paired validation.
The original cleaning policy excludes records flagged for negative future x, decreases in future x exceeding 0.1 m, absolute future heading above pi/2, or wrapped adjacent heading jumps above 1 radian. The new 54w source had 545,934 raw records and excluded 6,482; the new 149w source had 1,494,332 raw records and excluded 2,416. Cleaning manifests retain counts, reasons and source hashes.
## Provenance and license information
The image collections were identified in source records as `scene_pilot_2_dataset` and `ScenePilot-dataset`; trajectory labels were exported through an OpenVO pipeline and converted to the conversation format. Source machine paths in public provenance records are replaced by portable paths or basenames. Raw source exports, video downloads, model weights, training caches and training outputs are not included.
The source manifests did not provide a complete upstream image-source attribution list or a redistribution license. No new license is assigned by this dataset card. The uploader should supplement upstream attribution and applicable license information when available.
|