Towards-Realtime-MOT源代码学习之build_targets_thres()函数

该函数用于获取真实的置信度、目标框与id，其中目标框为投射到18*10大小特征图上的目标框

def build_targets_thres(target, anchor_wh, nA, nC, nGh, nGw):
    ID_THRESH = 0.5
    FG_THRESH = 0.5
    BG_THRESH = 0.4
    nB = len(target)  # number of images in batch
    assert(len(anchor_wh)==nA)

    tbox = torch.zeros(nB, nA, nGh, nGw, 4).cuda()  # batch size, anchors, grid size
    tconf = torch.LongTensor(nB, nA, nGh, nGw).fill_(0).cuda()
    tid = torch.LongTensor(nB, nA, nGh, nGw, 1).fill_(-1).cuda() 

传入进来的参数：nA=4，nC=1，nGh=10，nGw=18，anchor_wh=tensor([[ 2.6562,  7.9688],[ 3.7500, 11.2500],[ 5.3125, 13.1250],[10.6250, 10.0000]], device='cuda:0')，target=[tensor([         [0.0000e+00, 3.5300e+02, 9.1962e-02, 5.0159e-01, 5.3863e-02, 2.5566e-01],
        [0.0000e+00, 3.5500e+02, 6.0608e-01, 4.6625e-01, 5.1671e-02, 2.6615e-01],
        ...
        [0.0000e+00, 4.0100e+02, 4.5224e-02, 1.2969e-01, 4.6705e-02, 2.1583e-01],
        [0.0000e+00, 4.0200e+02, 1.0326e-01, 1.0715e-01, 3.7565e-02, 1.9927e-01]],
        device='cuda:0')]，target为一个list，里面装的Tensor的shape为[36, 6]，
nB = 1，表示batchsize
tbox.shape = torch.Size([1, 4, 10, 18, 4])
tconf.shape =  torch.Size([1, 4, 10, 18])
tid.shape = torch.Size([1, 4, 10, 18, 1])

for b in range(nB):
    t = target[b]
    t_id = t[:, 1].clone().long().cuda()
    t = t[:,[0,2,3,4,5]]
    nTb = len(t)  # number of targets
    if nTb == 0:
        continue

进入for循环之后
t=target[0]=tensor([[0.0000e+00, 3.5300e+02, 9.1962e-02, 5.0159e-01, 5.3863e-02, 2.5566e-01],
                               ...
                              [0.0000e+00, 4.0200e+02, 1.0326e-01, 1.0715e-01, 3.7565e-02, 1.9927e-01]],
       device='cuda:0')
t.shape = torch.Size([36, 6])
t_id = tensor([353, 355, 356, 357, 358, 359, 366, 367, 368, 369, 378, 379, 382, 386,
                      388, 394, 401, 402, 353, 355, 356, 357, 358, 359, 366, 367, 368, 369,
                      378, 379, 382, 386, 388, 394, 401, 402], device='cuda:0')
t = t[:,[0,2,3,4,5]]=tensor([ [0.0000, 0.0920, 0.5016, 0.0539, 0.2557],
                                          ...
                                          [0.0000, 0.1033, 0.1072, 0.0376, 0.1993]], device='cuda:0')
nTb = 36

gxy, gwh = t[: , 1:3].clone() , t[:, 3:5].clone()
gxy[:, 0] = gxy[:, 0] * nGw
gxy[:, 1] = gxy[:, 1] * nGh
gwh[:, 0] = gwh[:, 0] * nGw
gwh[:, 1] = gwh[:, 1] * nGh

gxy = t[: , 1:3].clone()=tensor([ [0.0920, 0.5016],
                                                 ...,
                                                 [0.1033, 0.1072]], device='cuda:0')
gwh= t[:, 3:5].clone()=tensor([ [0.0539, 0.2557],
                                                ...,
                                                [0.0376, 0.1993]], device='cuda:0')
gxy.shape=gwh.shape=torch.Size([36, 2])，gxy，gwh均乘上nGh和nGw之后：
gxy=tensor([ [ 1.6553,  5.0159],
                     ...,
                     [ 1.8587,  1.0715]], device='cuda:0')
gwh=tensor([ [0.9695, 2.5566],
                      ...,
                      [0.6762, 1.9927]], device='cuda:0')

gxy[:, 0] = torch.clamp(gxy[:, 0], min=0, max=nGw -1)
gxy[:, 1] = torch.clamp(gxy[:, 1], min=0, max=nGh -1)

此时gxy=tensor([[ 1.6553,  5.0159],
                            ...
                            [ 1.8587,  1.0715]], device='cuda:0')

gt_boxes = torch.cat([gxy, gwh], dim=1)  # Shape Ngx4 (xc, yc, w, h)
        
anchor_mesh = generate_anchor(nGh, nGw, anchor_wh)
anchor_list = anchor_mesh.permute(0,2,3,1).contiguous().view(-1, 4) 

gt_boxes=tensor([[ 1.6553,  5.0159,  0.9695,  2.5566],
                              ...,
                              [ 1.8587,  1.0715,  0.6762,  1.9927]], device='cuda:0')
anchor_mesh = tensor([[[[ 0.0000,  1.0000,  2.0000,  ..., 15.0000, 16.0000, 17.0000],
                                         ...,
                                         [ 0.0000,  1.0000,  2.0000,  ..., 15.0000, 16.0000, 17.0000]],
                                         ...,

                                         [[ 7.9688,  7.9688,  7.9688,  ...,  7.9688,  7.9688,  7.9688],
                                          ...,
                                          [ 7.9688,  7.9688,  7.9688,  ...,  7.9688,  7.9688,  7.9688]]],
                                          ...,
                                        [[[ 0.0000,  1.0000,  2.0000,  ..., 15.0000, 16.0000, 17.0000],
                                           ...,
                                           [ 0.0000,  1.0000,  2.0000,  ..., 15.0000, 16.0000, 17.0000]],
                                            ...,

                                          [[10.0000, 10.0000, 10.0000,  ..., 10.0000, 10.0000, 10.0000],
                                             ...,
                                           [10.0000, 10.0000, 10.0000,  ..., 10.0000, 10.0000, 10.0000]]]],
       device='cuda:0') 
anchor_mesh.shape= torch.Size([4, 4, 10, 18])
这里调用的generate_anchor函数如下：
anchor_list = tensor([[ 0.0000,  0.0000,  2.6562,  7.9688],
                                  ...,
                                  [17.0000,  9.0000, 10.6250, 10.0000]], device='cuda:0')
anchor_list.shape= torch.Size([720, 4])

iou_pdist = bbox_iou(anchor_list, gt_boxes)            # Shape (nA x nGh x nGw) x Ng
iou_max, max_gt_index = torch.max(iou_pdist, dim=1)    # Shape (nA x nGh x nGw), both

iou_map = iou_max.view(nA, nGh, nGw)       
gt_index_map = max_gt_index.view(nA, nGh, nGw)

这里调用了bbox_iou()函数计算iou，bbox_iou()函数如下：

iou_pdist = tensor([ [0.0016, 0.0000, 0.0465,  ..., 0.0000, 0.0857, 0.0000],
                                 ...,
                                [0.0000, 0.0000, 0.0000,  ..., 0.0000, 0.0000, 0.0000]],device='cuda:0')
iou_pdist.shape = torch.Size([720, 36])，[720]个anchor和[36]个gt_boxes（[4]表示目标框）分别求iou
iou_max.shape = 720，找出最大的720个iou值  
max_gt_index.shape=720，找出最大的720个iou值对应的索引    
iou_map.shape =   torch.Size([4, 10, 18])，此处的view函数用于重构张量的维度  
gt_index_map =torch.Size([4, 10, 18])

id_index = iou_map > ID_THRESH #将iou_map > ID_THRESH的位置置1，反之置0
fg_index = iou_map > FG_THRESH                                                    
bg_index = iou_map < BG_THRESH 
ign_index = (iou_map < FG_THRESH) * (iou_map > BG_THRESH)
#这一步是把fg_index作为索引，如fg_index中的一个元素为[0,0,1]，等于指定tconf中的[0,0,1]元素
tconf[b][fg_index] = 1
tconf[b][bg_index] = 0
tconf[b][ign_index] = -1

id_index.shape = fg_index.shape =bg_index.shape = ign_index.shape =torch.Size([4, 10, 18])
tconf=tensor([[[[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
                          ...,
                          [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]],

                          ...,

                         [[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
                           ...,
                          [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]]], device='cuda:0')
tconf.shape=torch.Size([1, 4, 10, 18])

gt_index = gt_index_map[fg_index]
gt_box_list = gt_boxes[gt_index]
gt_id_list = t_id[gt_index_map[id_index]]

gt_index = tensor([], device='cuda:0', dtype=torch.int64)
gt_box_list = tensor([], device='cuda:0', size=(0, 4))，得到符合条件的gt_boxes
gt_id_list = tensor([], device='cuda:0', dtype=torch.int64)，

if torch.sum(fg_index) > 0:
    tid[b][id_index] =  gt_id_list.unsqueeze(1)
    fg_anchor_list = anchor_list.view(nA, nGh, nGw, 4)[fg_index] 
    delta_target = encode_delta(gt_box_list, fg_anchor_list)
    tbox[b][fg_index] = delta_target

这里未进入if循环

return tconf, tbox, tid

最后将真实的置信度tconf、真实的目标框tbox、真实的id即tid返回