Parse27k Dataset

PARSE-27k基于在城市环境中由移动摄像机拍摄的8个不同长度的视频序列。DPM行人检测器已处理了序列的每15帧。我们用10个属性标签手动注释了生成的边界框，同时整理了检测器的误报。属性的选择是由机器人/汽车应用场景驱动的，并且具有与处理影响人的外观和形状的携带物品有关的多个视觉属性。 PARSE-27k有仔细的训练 (50%)，验证 (25%) 和测试 (25%) 拆分。这意味着我们只沿着序列边界分裂。此外，在同一天拍摄的序列可以是train-val或test。这避免了跨越分裂的高度相似的例子。与其他人属性数据集相比，PARSE-27k相对于姿势和裁剪具有相对较小的方差，因为它仅包含由行人检测器获得的行人边界框的作物。这与其他一些公开可用的数据集形成对比，这些数据集显示了各种不同的姿势和作物 (上半身镜头，只有脸，全身)。通过增加数据集大小和减少这种方差，我们希望提高模型质量。PARSE-27k is built upon 8 video sequences of varying lengths captured by a moving camera in urban environments. Every 15th frame of each sequence was processed by the DPM pedestrian detector. We manually annotated the resulting bounding boxes with 10 attribute labels, while also filtering out false positives from the detector. The selection of these attributes is driven by robotic/automotive application scenarios, and covers multiple visual attributes related to carried items that affect a pedestrian’s appearance and shape. PARSE-27k features a carefully designed train (50%), validation (25%), and test (25%) split. We only split the data along the boundaries of the video sequences. Additionally, sequences captured on the same day can be assigned to either the train-validation or test split, which avoids highly similar samples spanning different splits. Compared to other pedestrian attribute datasets, PARSE-27k has relatively low variance in terms of pedestrian pose and cropping, as it only contains crops of pedestrian bounding boxes obtained from the detector. This contrasts with some other publicly available datasets, which include a wide range of poses and cropping variations (e.g., upper-body shots, face-only crops, and full-body frames). By increasing the dataset scale and reducing this variance, we aim to improve model quality.

搜集汇总

背景与挑战

背景概述

Parse27k数据集基于城市移动摄像机视频，通过DPM检测器提取行人边界框并手动标注10个属性，专用于机器人或汽车应用场景。数据集进行了严格的训练、验证和测试拆分，以减少姿势和裁剪的方差，旨在提升模型质量。