zhreshold
About the reported performance
I find that in the camera-ready version of the SSD paper(http://link.springer.com/chapter/10.1007/978-3-319-46448-0_2/fulltext.html), the performance of "VOC07+12 trainval -> VOC07 test" is more than 74% and the reported performance here is 71.57%. Is there any difference between the training scheme/network structure here and that reported in the paper?
The original arXiv paper shows that it's 72.1. Which is almost identical. However, I checked that the author have updated some code, specifically modified some filter size and training hyper-parameters. So I guess he did achieved 2% improvement with those tweaks
How is the 71.57% achieved here?
I used the following commands: python train.py python evaluate.py --gpus 0,1,2,3 --batch-size 128 --epoch 100
and got ~69.9%.
I then downloaded the pretrained model and got 71.57% using: python evaluate.py --gpus 0,1,2,3 --batch-size 128 --epoch 0
Is it the same results you are getting?
In the latest SSD, the mAP is 77.2%, which is significantly better. There is also an tensorflow version here: https://github.com/balancap/SSD-Tensorflow
VOC07+12 trainval -> VOC07 test
I used this command: python train.py --gpus 1,2,3 --batch-size 64 --end-epoch 200
and then tested the 200th break-point and got a mAP about 73% is there a mxnet version can reach the profermance of latest SSD, ~77%?
