Few questions about accuracy in the paper

[skrish13]

Hi,

Q1. You had mentioned that the results of your method in the paper are from a weaker backbone than the baseline RetinaNet (which was resnet50?) [44]. But your paper mentions that your base detector has backbone of resnet50 (Section 3.1, Page 3). Could you please elaborate on the details?

Q2. You also mentioned [4] that initially you trained from scratch and later started using the ImageNet pretrained, is the paper's (RetinaNet and Full Approach) numbers from the scratch training or ImageNet pretrained?

Thanks for your time in advance :)

[eg4000]

Hi,

Q1. We originally used ResNet50 with a weaker FPN (less anchors and pyramid layers than RetinaNet)
Q2. We originally trained the "Full Approach" from scratch.

Regards,
Eran.

[skrish13]

Ah! Cool! Thanks for quick and clear response :)

[skrish13]

Just a quick follow up @eg4000 , the trained weights given in [9] is similar to the RetinaNet wrt network and is trained from ImageNet weights?

[eg4000]

yes