Cheng-Yang Fu

Hi @monocongo , I follow the COCO setting. If you can format your dataset in COCO format, you should be able to train it. But I haven't updated the code...

Hi, It looks like your priorbox setting is wrong. I assume you use Wei's VGG model. Because the stride, downsampling mechanism in ResNet is different from VGG. So, I change...

Hi, I don't think it's dataset reason. It looks like setting of PriorBox is wrong. Unfortunately, I am working on something urgent right now and cannot test this.

Hi, For the first question, I think the deconvolution layer is really useful for detecting small objects. According to the experimental results, DSSD get most improvement on small objects and...

Hi, @haibucuo0 , Just change the "start" field to "offset" in all PriorBox Layers. I try to make all the naming consistent with Wei's version. But there are still some...

Hi @jsjs0827 For the first question, I don't think deconvolutional layer will change preprocessing the input before operating it. Did you check the feature maps before feeding to deconvolutional layers?...

I didn't use the ReLU. You can try different values for the parameter "group". I think it will be helpful. Another thing I found is useful is adding 3x3 conv...

Hi @btbulxb , Obviously, there is something wrong. Can you upload the log let me check?

Hi @jinfagang, Sorry for the late reply. This is weird. Did you set the batch size as 1? The information in README is out of date. The current version should...

Not sure which part is wrong. But I think when trainning SSD321 with ResNet101, the "use_global_states" should be false. I turn that to True when training DSSD321.