TNL2K (Tracking by natural language)

自然语言跟踪（TNL2K）是为评估自然语言跟踪而构建的。 TNL2K 出现在：_x000D_
_x000D_
大规模：2,000 个序列，包含 1,244,340 帧，663 个单词，训练/测试分别为 1300 / 700 _x000D_
_x000D_
高质量：在每一帧中仔细检查的手动注释_x000D_
_x000D_
多模式：为每个序列提供视觉和语言注释_x000D_
_x000D_
Adversarial-samples：随机添加对抗性样本，用于对抗性攻防研究_x000D_
_x000D_
显着外观变化：包含行人 _x000D_ 的布/脸变化视频
_x000D_
异构：包含 RGB、热、卡通、合成数据_x000D_
_x000D_
多基线：BBox 跟踪、语言跟踪、联合 BBox 语言跟踪

Natural Language Tracking (TNL2K) is a dedicated dataset constructed for evaluating natural language tracking tasks.
Large-scale: The dataset contains 2,000 sequences, totaling 1,244,340 frames and 663 words, with training and test splits of 1300 and 700 respectively.
High-quality: All annotations are manually curated through careful inspection of each individual frame.
Multimodal: Visual and linguistic annotations are provided for every sequence.
Adversarial Samples: Adversarial samples are randomly incorporated into the dataset for adversarial attack and defense research.
Significant Appearance Variations: Features videos showcasing significant clothing and facial changes of pedestrians.
Heterogeneous: Covers diverse data modalities including RGB, thermal, cartoon, and synthetic data.
Multiple Baselines: Supports three core tracking baselines: Bounding Box (BBox) tracking, language tracking, and joint BBox-language tracking.

搜集汇总

背景与挑战

背景概述

TNL2K是一个用于评估自然语言跟踪的大规模高质量数据集，包含2000个序列和超过124万帧，提供视觉和语言注释，并涵盖多种数据模态和对抗性样本。该数据集由多所中国研究机构于2021年发布，支持目标跟踪和自然语言处理的交叉研究。