Regarding the quantitative target for the accuracy, we aim to achieve at least 85%~86% joint grading accuracy, which would be a 1~2% improvement from the current baseline design referring to CANet. The joint accuracy CANet is able to achieve 85.1% according to the paper, and the best joint accuracy we were able to replicate from its code was 84.5%. The table shown at the bottom right of our poster is our preliminary results which only our proposed grading algorithm was implemented, and we were able to achieve 85.2%. Currently, we are still experimenting and refining our proposed architecture by using different feature extraction and grading algorithms, but the main goal will remain to be enhancing the joint grading accuracy as much as possible.
To assess the accuracy, the main comparison would be on the joint accuracy performance of CANet. There is not much work that emphasised the joint grading performance of DR and DME, and to our knowledge, CANet was the best among similar existing works that address the joint accuracy. Moreover, our work would be an elaboration on CANet’s architecture, so the performance of our work would be best assessed by comparing it with CANet in terms of joint accuracy.
Do you have some quantitative target for accuracy ? How do you plan to assess the accuracy?
Thank you for your comment professor!
Regarding the quantitative target for the accuracy, we aim to achieve at least 85%~86% joint grading accuracy, which would be a 1~2% improvement from the current baseline design referring to CANet. The joint accuracy CANet is able to achieve 85.1% according to the paper, and the best joint accuracy we were able to replicate from its code was 84.5%. The table shown at the bottom right of our poster is our preliminary results which only our proposed grading algorithm was implemented, and we were able to achieve 85.2%. Currently, we are still experimenting and refining our proposed architecture by using different feature extraction and grading algorithms, but the main goal will remain to be enhancing the joint grading accuracy as much as possible.
To assess the accuracy, the main comparison would be on the joint accuracy performance of CANet. There is not much work that emphasised the joint grading performance of DR and DME, and to our knowledge, CANet was the best among similar existing works that address the joint accuracy. Moreover, our work would be an elaboration on CANet’s architecture, so the performance of our work would be best assessed by comparing it with CANet in terms of joint accuracy.