Best strategy for Object localisation without using pretrained model and weights.?

Question

我想从头开始训练模型，而无需使用任何预先训练的模型和权重以及20k图像数据集和4个对象坐标。您能建议我解决此问题的最佳方法吗？

Answer 1

Why won't you try any of the existing architectures - YOLO, SSD, Faster-RCNN to name a few (or their next-gens) and see what is the performance?
Of course, if you know in advance the properties of the objects in your data set this can be useful - eg for smaller objects, Faster-RCNN would probably yield better performance that SSD and YOLO.
Each one of them has available open source implementations in all leading DNN frameworks.