jiayulin/SeqGeo-VL-QA

SeqGeo-VL Q&A Dataset

Multiple-choice Q&A dataset for spatial reasoning and cross-view geo-localization evaluation.

Dataset Overview

This dataset contains 155,452 multiple-choice questions across 4 question types, based on 38,863 trajectories from the SeqGeo-VL dataset.

Question Types

1. text2sat: Given a text route description, select the correct satellite image
2. sat2text: Given a satellite image, select the correct route description
3. sat2video: Given a satellite image, select the correct street-view sequence
4. video2sat: Given a street-view sequence, select the correct satellite image

Each question has 4 multiple-choice options (1 correct + 3 spatially-diverse distractors).

Data Schema

Each entry contains:

• trajectory_id: Unique identifier
• question: Question text
• question_content: Input (text/image path/video frames)
• correct_answer: Ground truth answer
• choices: 4 multiple-choice options with labels
• context: Supplementary data (coverage bbox, trajectory coords, pixel coordinates)

Example Entry

text2sat, sat2text, video2sat (correct_answer is a string):

{
  "trajectory_id": "0",
  "question": "Which satellite image matches this trajectory description captured at street level?",
  "question_content": "The route begins at a busy urban intersection with brick commercial buildings...",
  "correct_answer": "satellite/2019_satellite/20/0.png",
  "choices": [
    {"label": "Option A", "modality_path": "satellite/2019_satellite/20/538.png"},
    {"label": "Option B", "modality_path": "satellite/2019_satellite/20/0.png"},
    {"label": "Option C", "modality_path": "satellite/2019_satellite/20/341.png"},
    {"label": "Option D", "modality_path": "satellite/2019_satellite/20/413.png"}
  ],
  "context": {...}
}

sat2video (correct_answer is a list of strings, choices have modality_paths):

{
  "trajectory_id": "0",
  "question": "Select the street-view image sequence that was captured along a trajectory shown in this satellite image:",
  "question_content": "satellite/2019_satellite/20/0.png",
  "correct_answer": [
    "street/2019_street/ROW/20190616.084458/000000281685.jpg",
    "street/2019_street/ROW/20190616.084458/000000289759.jpg",
    "street/2019_street/ROW/20190616.084458/000000297861.jpg"
  ],
  "choices": [
    {
      "label": "Option A",
      "modality_paths": ["street/2019_street/ROW/20190616.085039/000010782750.jpg", ...]
    },
    {
      "label": "Option B",
      "modality_paths": ["street/2019_street/ROW/20190616.085039/000012786631.jpg", ...]
    }
  ],
  "context": {...}
}

Key Features

• Spatial Diversity: Distractors are >5km away (haversine distance)
• Randomized Choices: Correct answer position varies per entry
• Template Variety: Multiple question phrasings per type
• Rich Context: Geographic and pixel coordinates included
• Inference-Ready: All content inline, no external lookups needed

File Structure

text2sat/
├── train.json  
└── val.json    

sat2text/
├── train.json  
└── val.json    

sat2video/
├── train.json  
└── val.json    

video2sat/
├── train.json  
└── val.json    

Loading the Dataset

from datasets import load_dataset

# Load a specific config (text2sat, sat2text, sat2video, or video2sat)
ds = load_dataset("jiayulin/SeqGeo-VL-QA", "text2sat")

# Access train and validation splits
train_data = ds["train"]
val_data = ds["validation"]

# Or load directly
ds_text2sat = load_dataset("jiayulin/SeqGeo-VL-QA", "text2sat", split="train")

Usage Examples

For text2sat, sat2text, video2sat (single correct answer):

ds = load_dataset("jiayulin/SeqGeo-VL-QA", "text2sat")

for entry in ds["train"]:
    question = entry["question"]
    question_content = entry["question_content"]
    choices = entry["choices"]  # List of choice objects
    correct_answer = entry["correct_answer"]  # Single string
    
    # Access modality_path from each choice
    options = [c["modality_path"] for c in choices]

For sat2video (list of correct answers):

ds = load_dataset("jiayulin/SeqGeo-VL-QA", "sat2video")

for entry in ds["train"]:
    question = entry["question"]
    question_content = entry["question_content"]
    choices = entry["choices"]  # List of choice objects
    correct_answer = entry["correct_answer"]  # List of strings
    
    # Access modality_paths (plural) from each choice
    options = [c["modality_paths"] for c in choices]  # Each is a list

Citation

If you use SeqGeo-VL in your research, please cite both TrajLoc, which introduces the vision-language annotations, and SeqGeo, which provides the original imagery and trajectory data.

TrajLoc and SeqGeo-VL

@inproceedings{gao2026trajloc,
  title     = {Trajectory-aware Cross-view Geo-localization with Sequential Observations},
  author    = {Gao, Tianyi and Lin, Jiayu and Beaulieu, Danielle and Jacobs, Nathan},
  booktitle = {European Conference on Computer Vision},
  year      = {2026}
}

Original SeqGeo Dataset

@inproceedings{zhang2023seqgeo,
  title     = {Cross-View Image Sequence Geo-Localization},
  author    = {Zhang, Xiaohan and Sultani, Waqas and Wshah, Safwan},
  booktitle = {Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
  pages     = {2914--2923},
  year      = {2023}
}

Source

Based on the SeqGeo-VL dataset. For more information, see the original SeqGeo-VL resources:

• 🌐 TrajLoc Project Page
• 📄 TrajLoc Paper
• 💻 GitHub Repository
• 🤗 Pretrained Weights