Actually we are using faster_rcnn_inception_resnet_v2_atrous_coco pre-trained models, to train over our own dataset images, but we want to improvement our object detection.
How can I train an model from scratch, for example, using inception v3 or v4 to object detection using a COCO dataset? I wanna generate a pre-trained model called faster_rcnn_inception_v3_coco like others described at https://github.com/tensorflow/models/blob/master/research/object_detection/g3doc/detection_model_zoo.md
You can manually replace the base network, e.g. with these tensorflow-slim models pretrained on ImageNet: https://github.com/tensorflow/models/tree/master/research/slim (Readme: Pre-trained Models)
I took the weights and the model of InceptionV4 and used it as a base network for SSD (Single Shot Multibox Detector). However, it requires some knowledge of tensorflow.
If you don't have time or lack the knowledge, you might want to stick with the high level APIs. You could try out the NASnet model.
