Egocentric Video Action Planning Benchmark with 741 Samples

EgoPlan-Bench provides 741 samples for evaluating models on egocentric video understanding and action planning. Each sample contains a video file, a question about the next action, a task goal, and four multiple-choice options with a correct answer. The dataset is designed as a benchmark for tasks involving video and textual reasoning.

Use Cases

• Train multimodal models to predict the answer from video and question inputs for action planning.
• Benchmark video-language models on the task_goal and question fields to assess reasoning in egocentric scenarios.
• Analyze model performance across the 4 options per sample to identify common failure modes in action selection.

Strengths

• Contains 741 curated benchmark samples for a specific research task.
• Includes multiple data modalities: video files and structured text fields like question and task_goal.
• Provides a clear evaluation framework with defined correct answers for each sample.

Limitations

• Modest size of 741 samples may limit training complex models from scratch.
• Lacks detailed column specifications and sample data for full structural transparency.
• Potential for narrow scope focused on a specific benchmark, limiting generalizability.

Provenance

Source

Hugging Face dataset authored by advaitgupta.

Collection Method

Curated for the EgoPlan-Bench benchmark, methodology details unknown.

Freshness

Last updated on 2025-11-11.