引导式测验 · 中等 等级

开放词汇对象检测测验

回顾文本条件区域检测、图像文本预训练、新颖的类别提示、本地化以及评估生成的注释的需要。

相关引导路径开放词汇目标检测
问题 1 的 8

开放词汇目标检测的定义能力是什么？

预览每个问题8 问题

当您参加测验时，就会出现答案和解释。

  1. 开放词汇目标检测的定义能力是什么？
    • 仅检测训练期间用框看到的类别
    • 比所有其他探测器运行得更快
    • 检测任意文本描述的对象，包括看不见的类别
    • 仅适用于灰度图像
  2. In a common embedding-based open-vocabulary detector, how are candidate regions matched to user-supplied labels?
    • By using only the region's average brightness
    • By comparing visual-region features with text embeddings in a compatible representation space
    • By choosing labels alphabetically without visual features
    • By restricting every model to the original COCO class list
  3. 哪种预训练资源最能实现开放词汇检测？
    • 用于学习视觉语言空间的大规模图像文本对
    • 一小套贴有完美标签的盒子
    • 录音
    • 手写数字数据集
  4. What replaces fixed per-class classifier weights in a common embedding-based open-vocabulary design?
    • The camera's physical image sensor
    • The bounding-box coordinate system
    • All visual features with random file identifiers
    • Representations of the supplied text labels used for similarity-based classification
  5. 以下哪一个是开放词汇或基础检测模型的示例？
    • 单独 ResNet-50
    • 恐龙接地
    • 一个简单的 k-means 聚类脚本
    • 电子表格宏
  6. What does an open-vocabulary detector's ability to use a new text category imply?
    • It guarantees accurate detection of every nameable object
    • The category must have been absent from every pretraining image and caption
    • It can attempt localization through learned vision-language associations, with accuracy requiring evaluation
    • It needs no visual features or bounding-box output
  7. What output distinguishes text-conditioned object detection from simply saying that an image contains a named category?
    • Only a whole-image tag with no object location
    • Only the text prompt's word count
    • Only the image's dimensions and filename
    • Locations such as bounding boxes for the matching objects
  8. Before using open-vocabulary predictions to auto-label a new domain, what should a dataset team check?
    • The boxes and labels under representative prompts, thresholds, and domain conditions
    • Assume text-conditioned predictions are automatically ground truth
    • Verify only that the label strings contain no spelling errors
    • Discard all visual inspection because image-text pretraining guarantees accuracy