Done

Start using IntelliJ to take advantage of debugger to read source code.
Find a good repo that has most thing structured and well-commented. I cloned this repo and made changes to the FasterRCNN project, the main function should mostly work.
- Most of fitting work is import-related. Relative path doesn’t work in my project, but I’m worry about importing the whole script will slow down the code running.
- In train_utils/coco_eval, import torch.six_ to call a string_class doesn’t work because unable to import `torch._six`. This could because of torch version, what i’m currently working on is “2.0.1”.
- Some highlights of this Repo:
  - COCO metrics like mAP (Mean Average Precision) has been nicely set up, so I don’t need to spend time recreating the wheels. We should be able quantitatively evaluate our model not only by how the overall detection works, but also by its performance on large/median/small ground truth boxes. These categorized performance report probably will help us to identify whether the “receiptive filed” guess stands a point or not.
  - The whole structure can be easily edited if we have some new ideas of modifying the structure.
Read a paper about generating oriented boxes. Basically it’s replacing the RPN part in Faster RCNN with a new network that propose oriented boxes.
- More than just generating 4 box regression parameters for each box, this network generates anotehr rotation angle parameter \(\theta\) to calculate loss.
  - It has a coarse box scores that responsible and then refine the parameters to generate proposal regions.
- This network is similar to RNP in terms of positive/ negative sampling to calculate loss.
- Since is is like a “better-RPN”, this network can be easily benchmark to Fast RCNN for training.
- Obviously the RoI’s are better, so does the mAP.

Todos

Reimplement split_data function
Reimplement a function to modify xml files, so we could have a result as a starting point for the new model.
- Probably use an argument to indicate if I want to do oriented proposals.
- Really need to add is_crowd attribute.
- Need to add get_width and get_height method in my_dataset.
Skimmed at Samia’s project, it’s a decomposition + a GAN model, need to look into the details to get that run with our data.
- Could this be an easier replacement?