MVVLP: Multi-View Dataset for Vision-and-Language Parking

MVVLP includes a multi-view image dataset collected in real parking lots, a parking instruction dataset generated by large language model, an evaluation platform for evaluating AVP algorithms, and implementations of several reinforcement learning methods.It focuses on autonomous valet parking, supporting: Multi-view Perception: Leverages synchronized camera feeds from various angles to provide a comprehensive view of the parking environment.️ Natural Language Instructions: Uses LLM-generated parking commands to simulate human-to-vehicle interactions and enable instruction-following capabilities.

Simulation & Evaluation: Offers a unified platform to simulate valet parking scenarios and quantitatively evaluate algorithmic performance across benchmarks. Reinforcement Learning Frameworks: Includes baseline implementations (e.g., DQN, PPO, SAC) tailored to the AVP task for training and benchmarking.️ Real-world dataset integration: Supports training and testing using data collected from real parking lots to bridge the gap between simulation and reality.MVVLP includes a multi-view image dataset collected in real parking lots, a parking instruction dataset generated by large language model, an evaluation platform for evaluating AVP algorithms, and implementations of several reinforcement learning methods.The multi-view image dataset contains 4 parking structures, 144 shooting points, 4 different perspectives, 17 collocation times and 2492 privacy-processed images.Based on the multi-view images, 144 parking spaces, 2736 parking space inherent attributes and 1178 parking states are manually annotated.The natural language instruction dataset contains 5 distinct instruction styles, 150 sets of semantic representations, a total of 750 demand-driven instructions and another 250 random instructions.A detailed description of the dataset can be found in the README.md.