Six paragraph including introduction, related works, methodology, training, experiments, and conclusion rather than five. Is it because this paper proposed a new network model(RefineDet) so that they must describe the training details?
Training and Inference
The organization of this part should be learned. Firstly, Data Augmentation and Backbone Network must be presented. Then we should describe our contributions in detail. Generally, we can start with some common improvements such as loss function, optimization, and activation. Finally, the inference should be stated(If it's not our contributions, then we needn't stat particular detail. And we needn't show the results either).
Experiments
We should test our model on as many open datasets as possible and compare with those state-of-the-art methods. The comparison comes from runtime performance and accuracy. In addition, we should test the ablation and transformation of our model to demonstrate its effectiveness. And the main contribution or changes we performed on the network must be tested in experiments.
Best trade-off between accuracy and speed.
