detection results have deviation with coco dataset

[haibucuo0]

Hi, I want to know why I got such detection results in following figures. Figure 1 and Figure 2 show the results using trained dssd and ssd model(coco dataset) respectly. Rectangular boxes drawn according to detection results are not very accurate frame to detect objects in Figure1 and Figure2. Figure 3 is the result using trained ssd model(voc0712 dataset) that weiliu shared, and the Rectangular boxes drawn are very accurate. May the internet speed is not good， so I can't upload Figure to github. To deal with problem, I have send an email to your mailbox([email protected]), my email is [email protected]. Thank you. chen

[chengyangfu]

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 the parameters of Priorboxs too.

[haibucuo0]

Hi, Sorry, maybe the description of the sentence"Figure 1 and Figure 2 show the results using trained dssd and ssd model(coco dataset) respectly" is not clear. The trained dssd and ssd model(coco dataset) I used were shared from you. Are there any other reasons? Thank you.

[chengyangfu]

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.

[haibucuo0]

Hi, Do you mean I need to modify the PriorBox layer in deploy.prototxt？