wei-yuma / multitrident

I combine trident net and refinedet to detect objects of different scales at the same time in one forward propagation

multitrident(updating)

This is an object detection project. What I want to do is to let detector can detect objects of different scales through different branches, so I can detect and show detected objects of different scales in different detect heads. It dosen't need judge scale of objects and show them, because the different branches of network has learn what it small object、what is middle object and what is large object.

network

performance on voc0712

img_size = 320*320

class	aero	bike	bird	boat	bottle	bus	car	cat	chair	cow	table	dog	horse	mbike	person	plant	sheep	sofa	train	Tv
RefineDet AP	83.9	85.4	81.4	75.5	60.2	86.4	88.1	89.1	62.7	83.9	77.0	85.4	87.1	86.7	82.6	55.3	82.7	78.5	88.1	79.4
multitrident AP	84.5	84.4	76.9	70.0	63.7	85.0	87.9	87.5	62.2	80.4	67.7	83.4	87.4	85.2	81.6	50.9	80.9	76.0	84.6	77.4

demo

all output

only one branch output

all output small object detect branch output middle object detect branch output large object detect branch output

multitrident's Issues

Should it be conv weight sharing for the trident net?

I read the tridentnet paper briefly and wondered if it should be using the same conv weight for different dilated conv kernels

你好，问下训练相关的问题

1.这个网络为什么 每隔一个epoch，损失才明显降低?
2.为什么 我设置 num_workers=4,8或其他数 损失就不会降低？
3.我用DistributedDataParallel 总是会报错：set 'find_unused_parameters=True` to torch.nn.parallel.DistributedDataParallel?
都是些好奇怪的错误。

谢谢!

large small medium size obj how to work ?

"Legacy autograd function with non-static forward method is deprecated. "

"Legacy autograd function with non-static forward method is deprecated. "
RuntimeError: Legacy autograd function with non-static forward method is deprecated. Please use new-style autograd function with static forward method. (Example: https://pytorch.org/docs/stable/autograd.html#torch.autograd.Function)

Any idea how to solve it?

ROIAlign

I noticed that you wrote ROIAlign in the regionpooling_refinedet code, but these codes are commented. I would like to ask how to add the ROIAlign algorithm to RefineDet,please~

I was training a custom dataset, but after view iteration i started getting NAN values. What could be the possible reason ?

iter 30 || ARM_L: nan ARM_C: nan TRM_s_L: nan TRM_s_C: nan TRM_m_L: 0.0000 TRM_m_C: 0.0000 TRM_b_L: 0.0000 TRM_b_C: 0.000

what is the meaning of extra_layers[1::2]?Thank for your help.