The faster-rcnn from oymiss

faster-rcnn's People

Contributors

Stargazers

Watchers

faster-rcnn's Issues

Questions about the regression of delta

Hi, thanks for useful work. I'm not sure about the calculation of the delta regression.
In the rcnn_loss_layer function, I think the shape of delta is (p, c, 4) and p is the sum of all proposals in all batches.
For the i-th batch, find the index of a particular class, then accesses delta using the index

faster-rcnn/layers/rcnn_loss_layer.py

Lines 12 to 43 in f219ebb

 def rcnn_loss_layer(score, delta, label_list, fg_delta_list, index_list): 

 # score (p, c) 

 # delta (p, c, 4) 

 # label_list list(n, l) sum(l) == k 

 # fg_delta_list list(n, l1, 4) 

 # index_list List[Tuple[Tensor, Tensor]] 

 # return loss() 

 p, num_class = score.shape[:2] 

 reg_loss, cls_loss = torch.tensor(0.), torch.tensor(0.) 

 n = len(index_list) 

 # labels = torch.cat(label_list) 

 for i in range(n): 

 label = label_list[i] 

 fg_i, bg_i = index_list[i] 

 index = torch.cat((fg_i, bg_i)) 

 # (k,) 

 cls_loss += F.cross_entropy(score[index], label) 

 # (l1, 4) 

 fg_delta_i = fg_delta_list[i] 

 for j in range(1, num_class + 1): 

 index_j = torch.where(label == j) 

 x = fg_i[index_j] 

 if x.shape[0] == 0: 

 continue 

 pred_delta_j = delta[x, j] 

 gt_delta_j = fg_delta_i[index_j] 

 if pred_delta_j.shape[0] > 0: 

 reg_loss += loss_tools.smooth_l1_loss(pred_delta_j, gt_delta_j) 

 return reg_loss + cls_loss

x = fg_i[index_j]
pred_delta_j = delta[x, j]

But I think the index of x here is to indicate the corresponding proposals in the batch of i.
In other words, since delta has a length of P that concatenates all the batches, it does not access the batch number i using x. I think I need to stagger the idx to account for batches to access the correct delta, but what about that?

Recommend Projects

oymiss / faster-rcnn Goto Github PK

faster-rcnn's People

Contributors

Stargazers

Watchers

faster-rcnn's Issues

Questions about the regression of delta

Recommend Projects

React

Vue.js

Typescript

TensorFlow

Django

Laravel

D3

Recommend Topics

javascript

web

server

Machine learning

Visualization

Game

Recommend Org

Facebook

Microsoft

Google

Alibaba

D3

Tencent

	def rcnn_loss_layer(score, delta, label_list, fg_delta_list, index_list):
	# score (p, c)
	# delta (p, c, 4)
	# label_list list(n, l) sum(l) == k
	# fg_delta_list list(n, l1, 4)
	# index_list List[Tuple[Tensor, Tensor]]
	# return loss()
	p, num_class = score.shape[:2]
	reg_loss, cls_loss = torch.tensor(0.), torch.tensor(0.)
	n = len(index_list)

	# labels = torch.cat(label_list)
	for i in range(n):
	label = label_list[i]
	fg_i, bg_i = index_list[i]
	index = torch.cat((fg_i, bg_i))
	# (k,)
	cls_loss += F.cross_entropy(score[index], label)

	# (l1, 4)
	fg_delta_i = fg_delta_list[i]
	for j in range(1, num_class + 1):
	index_j = torch.where(label == j)
	x = fg_i[index_j]
	if x.shape[0] == 0:
	continue
	pred_delta_j = delta[x, j]
	gt_delta_j = fg_delta_i[index_j]
	if pred_delta_j.shape[0] > 0:
	reg_loss += loss_tools.smooth_l1_loss(pred_delta_j, gt_delta_j)

	return reg_loss + cls_loss