metadata
configs:
- config_name: direction_only_close_ended
data_files:
- split: val
path: direction_only_close_ended.json
- config_name: direction_only_open_ended
data_files:
- split: val
path: direction_only_open_ended.json
- config_name: direction_obj_in_q_close_ended
data_files:
- split: val
path: direction_obj_in_q_close_ended.json
- config_name: direction_obj_in_q_open_ended
data_files:
- split: val
path: direction_obj_in_q_open_ended.json
- config_name: direction_obj_in_a_close_ended
data_files:
- split: val
path: direction_obj_in_a_close_ended.json
- config_name: direction_obj_in_a_open_ended
data_files:
- split: val
path: direction_obj_in_a_open_ended.json
- config_name: direction_obj_in_qa_close_ended
data_files:
- split: val
path: direction_obj_in_qa_close_ended.json
- config_name: direction_obj_in_qa_open_ended
data_files:
- split: val
path: direction_obj_in_qa_open_ended.json
- config_name: object_recognition_close_ended
data_files:
- split: val
path: object_recognition_close_ended.json
- config_name: object_recognition_open_ended
data_files:
- split: val
path: object_recognition_open_ended.json
license: mit
task_categories:
- video-classification
- question-answering
language:
- en
tags:
- video
- spatial-reasoning
- direction
- VideoLLM
pretty_name: E2E Real Object Direction
size_categories:
- n<1K
E2E Real Object Direction
A video-based benchmark for evaluating VideoLLMs' directional reasoning and object recognition on real-world objects.
Conditions
| Condition | Question | Answer | Purpose |
|---|---|---|---|
direction_only |
"In which direction is the object moving?" | "Up" | Baseline direction recognition |
direction_obj_in_q |
"In which direction is the car moving?" | "Up" | Does naming the object help? |
direction_obj_in_a |
"In which direction is the object moving?" | "The car is moving up" | Object grounding in answer |
direction_obj_in_qa |
"In which direction is the car moving?" | "The car is moving up" | Full object grounding |
object_recognition |
"What is the object moving in this video?" | "Car" | Object identification only |
Each condition has both close_ended (MCQ) and open_ended (free-form) variants.
Usage
from datasets import load_dataset
ds = load_dataset("YOUR_HF_ID/E2E_real_object", name="direction_only_close_ended", split="val")
Data Format
Direction tasks (close_ended)
{
"id": 0,
"video": "up/horse.mp4",
"category": "up",
"question": "In which direction is the object moving in this video?",
"options": ["Up", "Down", "Left", "Right"],
"answer": "A"
}
Object recognition (close_ended)
{
"id": 0,
"video": "up/horse.mp4",
"category": "up",
"question": "What is the object moving in this video?",
"options": ["Horse", "Car", "Dog", "Laptop"],
"answer": "A"
}
Video Structure
E2E_real_object/
├── up/
│ ├── car.mp4
│ ├── dog.mp4
│ └── ...
├── down/
├── left/
└── right/
Objects
bicycle, car, dog, bed, basketball, bench, sophia, chair, sofa, horse, laptop, duck, sphere