ReaMOT Challenge
收藏资源简介:
ReaMOT Challenge是一个基于12个数据集构建的推理式多目标跟踪基准,旨在推动基于推理的多目标跟踪任务的研究。该数据集包含1156条具有推理特性的语言指令,423,359个图像-语言对和869个不同场景,分为三个推理难度等级:简单、中等和困难。数据集的创建过程包括手动预选、GPT辅助标注和手动标注与复查。ReaMOT Challenge旨在解决在复杂语言指令下进行多目标跟踪的挑战,并评估跟踪模型的推理能力。
The ReaMOT Challenge is a reasoning-based multi-object tracking benchmark constructed based on 12 datasets, aiming to advance research on reasoning-based multi-object tracking tasks. This dataset contains 1,156 reasoning-specific language instructions, 423,359 image-language pairs, and 869 distinct scenarios, which are categorized into three reasoning difficulty levels: easy, medium, and hard. The dataset creation process includes manual pre-selection, GPT-assisted annotation, as well as manual annotation and review. The ReaMOT Challenge aims to address the challenges of multi-object tracking under complex language instructions and evaluate the reasoning capabilities of tracking models.

- 1ReaMOT: A Benchmark and Framework for Reasoning-based Multi-Object Tracking华中科技大学 · 2025年



