chengyangfu
resnet questions
Hi, I have a question and a request: Q:Without deconvolution layers,do you think it is helpful to detect the small objects just using Resnet model instead of VGG model? R:Could you share your trained voc ssd and dssd model, coco ssd and dssd model?At home, these can not be downloaded. My email is [email protected]. Thank you. Chen
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 medium objects. Current SSD(VGG) is already pretty good at detecting medium and large objects. Just changing base network from VGG to ResNet only brings very few improvement.
I already put all the files in the google drive. Unfortunately, the file is too big, I don't think it can be sent through e-mail.
Hi, 1.When I used the trained models of coco to test image directly, there are some errors： [libprotobuf ERROR google/protobuf/text_format.cc:245] Error parsing text-format caffe.NetParameter: 13235:10: Message type "caffe.PriorBoxParameter" has no field named "start". F0402 20:19:03.518262 3219 upgrade_proto.cpp:88] Check failed: ReadProtoFromTextFile(param_file, param) Failed to parse NetParameter file: models/ResNet-101/deploy.prototxt *** Check failure stack trace: *** @ 0x7f2b876975cd google::LogMessage::Fail() @ 0x7f2b87699433 google::LogMessage::SendToLog() @ 0x7f2b8769715b google::LogMessage::Flush() @ 0x7f2b87699e1e google::LogMessageFatal::~LogMessageFatal() @ 0x7f2b87c95021 caffe::ReadNetParamsFromTextFileOrDie() @ 0x7f2b87c63b25 caffe::Net<>::Net() @ 0x40944a Detector::Detector() @ 0x405182 main @ 0x7f2b85d71830 __libc_start_main @ 0x4060c9 _start @ (nil) (unknown) 2.Could you share the trained models of ssd and dssd based on voc dataset?I only downloaded coco trained models.Thanks
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 mistakes. :(
For the second question, I am working on PyTorch version of SSD/DSSD. Not sure I could provide the models soon.