I have a dataset which has satellite images. So dataset is quite different from usual image datasets used for object detection. I trained ssd_mobilenet_v1_pets model, but trained model performs really bad.
Does it mean that ssd_mobilenet_v1_pets is not a good candidate for satellite imagery? And which one of all other available models in TF object detection would be a better performer in my case?
The best choice of model depends on your exact requirements and deployment environment. SSD Mobilenet (and other SSD models) perform inference very quickly, but with less accuracy. They are well suited to cases where fast/realtime inference is desirable or situations where computing power is limited (ie. Mobile phones or IoT). In contrast, Faster RCNN or RFCN models will yield more accurate results, but run slower.
Consider trying the Faster RCNN Resnet 101 models. If you want more details check out the model zoo or this paper on the speed accuracy trade-offs of object detection architectures.