We can't get the mAP of baseline in paper

[wcg5262]

We use pytorch.faster-RCNN to try to reproduce the baseline on VOC in the paper. We found that the baseline mAP described in the paper was not reached. Compared with joint training, it can increase 20 to nearly 40 percentage points with only fine-tuning, especially under the premise that nove class samples have never been seen in the first training phase. Can you specifically describe how you set the experimental data and parameters for fine-tune and ft-full.

[zr526799544]

We use pytorch.faster-RCNN to try to reproduce the baseline on VOC in the paper. We found that the baseline mAP described in the paper was not reached. Compared with joint training, it can increase 20 to nearly 40 percentage points with only fine-tuning, especially under the premise that nove class samples have never been seen in the first training phase. Can you specifically describe how you set the experimental data and parameters for fine-tune and ft-full.

Hello. could you mind tell me the way you train the model and the way you build the datasets?