Argoverse-HD

Argoverse-HD是一个关于流式对象检测的数据集，包括实时对象检测、视频对象检测、跟踪和短期预测。该数据集包含来自Argoverse 1.1的视频数据，总共70,000个图像帧和130万个边界框，带有MS COCO风格的注释和轨道id，分辨率1920年x 1200 @ 30 FPS。这些注释与COCO向后兼容，因此研究人员可以直接评估此数据集上的COCO预训练模型，以估计模型的效率或跨数据集的泛化能力。
Argoverse-HD是用于流量感测挑战的数据集，其中包括两个轨道:
单一检测 (实时对象检测): 在这条赛道上，竞争对手将开发单帧对象检测器，类似于COCO和LVIS挑战。关键的区别在于，评估将通过流的准确性对延迟进行评分。全栈: 在这条轨道上，方法是不受限制的。但是，最有可能的是，跟踪和预测将用于补偿检测器延迟。
默认情况下，所有提交的延迟都由V100 GPU的官方工具包测量。

Argoverse-HD is a dataset dedicated to streaming object detection, encompassing real-time object detection, video object detection, tracking and short-term prediction. It comprises video data sourced from Argoverse 1.1, with a total of 70,000 image frames and 1.3 million bounding boxes, annotated in MS COCO-style with track IDs, at a resolution of 1920 × 1200 @ 30 FPS. These annotations are backward-compatible with COCO, allowing researchers to directly evaluate COCO-pre-trained models on this dataset to assess model efficiency or cross-dataset generalization ability.

Argoverse-HD serves as the dataset for the Traffic Sensing Challenge, which features two tracks:
1. Single Detection (Real-Time Object Detection): In this track, participants will develop single-frame object detectors analogous to those employed in the COCO and LVIS challenges. The key difference is that evaluation will score both streaming accuracy and latency.
2. Full Stack: There are no methodological restrictions for this track. However, it is most prevalent to utilize tracking and prediction modules to compensate for detector latency.

By default, the latency of all submissions is measured using the official toolkit on a V100 GPU.

搜集汇总

背景与挑战

背景概述

Argoverse-HD是一个专注于流式对象检测的数据集，包含70,000个图像帧和130万个边界框，采用MS COCO风格的注释和轨道ID，适用于实时对象检测、视频对象检测、跟踪和短期预测任务。该数据集支持COCO预训练模型的评估，并用于流量感测挑战，包含单一检测和全栈两个赛道。