Is there any other hidden settings or tricks for training

[AlexSunNik]

I followed your paper and tf code and reimplemented your model in PyTorch and trained from scratch but can only get 77.9 map for VOC2007

[AlexSunNik]

Also, I'm wondering how you're achieving 80.0 mAP on VOC12 Val using only VOC12 train as the training dataset? The results were published on your paper. Would you mind sharing the optimization settings? Also, have you finetuned your model or the results are collected right after training from scratch.

[dvornikita]

Hello,

We were trying to be clear in the paper on the implementation details and to be sure there is no ambiguity left we have released the code. There are no intentional hidden settings and tricks. You must be able to reproduce the reported results with the code in this repository.

When evaluating on VOC12, we train the model on the union of VOC07-train and VOC12-train.
I hope that helps.

Best,
Nikita

[AlexSunNik]

Hi thanks for your reply! I'm just wondering, with VOC2007 trainval + VOC2012 trainval, you're able to get a 75.4 map on VOC2012 test. With VOC2007 train and VOC2012 train only, how to get a 80.0 map tested on VOC2012 val? It's the Table 3 in your paper I'm referring to. I ran the experiment and can only get 73.5 map with your provided optimization settings to train the BlitzNet 300 model.

[dvornikita]

Sorry, I may have misunderstood you before.
If you mean the Results in Table 3, we have used a set that we call VOC12trainseg-aug with extra segmentation annotations. You can see the details and the corresponding reference to the annotations in Section 4.1 of the paper.