Feature backbone采用DLA,输入维度为3×H×W的RGB图,得到维度D×h×w的特征图F,然后将特征图送入几个轻量级regression heads,2D bouding boxes的中心特征图用下面的模块得到:
其中AN是Attentive Normalization.用公式表示:
类似的, 2D和3D bouding boxes的中心之间的offset用公式表示:
深度用以下两个公式表示:
其中Z是预测的深度值,delta是标准差.
3D bouding boxes的维度用以下公式表示:
- Monocular Auxiliary Detection Learning Contextsmonocular auxiliary detection learning experience learning contexts buffers generalizable monocular mononerf learning surveillance detection learning maritime contrastive supervised detection learning detection learning anomaly survey auxiliary contexts stability-plasticity plasticity achieving auxiliary monocular