论文标题：End-to-End Object Detection with Transformers
论文官方地址：https://ai.facebook.com/research/publications/end-to-end-object-detection-with-transformers
个人整理的PPT（可编辑），下载地址：DETR学习分享.pptx
一、什么是匈牙利算法
匈牙利匹配算法，是一种典型的一对一的配对算法。一对一的匹配，与之对应的，就是一对多，或多对一的配对算法。也称作 no anchor 操作。
对于Anchor based 、 Anchor free 和 no anchor 的辨析，可以参考这里：【AI面试】Anchor based 、 Anchor free 和 no anchor 的辨析
https://qianlingjun.blog.csdn.net/article/details/129339036在目标检测算法中，预测结果与gt标注结果多对一的典型案例，就是anchor based时候，对推荐的特征框是很多的，但是图像中的标记目标是很少的。对于阳性和阴性案例的分配时候，就采用了IOU 的方式来判别。
- 大于0.7的就是positive，
- 小于0.3的就是negative。
- 即便如此，positive的数量还是相比于标记框数量还是多的，此时的状态就是多对一的。
（上面对于阳性样本和阴性样本的划分，是一个简略的介绍，更详细的可以参考这里：【AI面试】YOLO 如何通过 k-means 得到 anchor boxes的？Yolo、SSD 和 faster rcnn 的正负样本定义）
在匈牙利匹配算法组，就是要找到一种一对一的组合方式下，目标是最优的。这里举一个案例：
一个农场主有10名工人，他们分别都有自己擅长的事情，现在需要做5件不同的事情，一个人只能干一件事情。如何分配能够使得最后的工作效率最高呢？
最简单的方式，就是采用遍历的方式，把所有的可能都计算一遍。最后把工作效率最高的一个组合留下来。此时，这种组合就是最优的匹配。
匈牙利匹配算法就是采用这种方式进行的。步骤如下：
- 先定义一个目标任务，怎么判断是最优匹配？比如这里就是误差最小；
- 列举所有的可能，最终找到误差最小的那个组合，就是一一匹配的最优组合形式了。
二、代码
match.py 部分完整代码，如下：
# Copyright (c) Facebook, Inc. and its affiliates. All Rights Reserved
"""
Modules to compute the matching cost and solve the corresponding LSAP.
"""
import torch
from scipy.optimize import linear_sum_assignment
from torch import nn
from util.box_ops import box_cxcywh_to_xyxy, generalized_box_iou
class HungarianMatcher(nn.Module):
"""This class computes an assignment between the targets and the predictions of the network
For efficiency reasons, the targets don't include the no_object. Because of this, in general,
there are more predictions than targets. In this case, we do a 1-to-1 matching of the best predictions,
while the others are un-matched (and thus treated as non-objects).
"""
def __init__(self, cost_class: float = 1, cost_bbox: float = 1, cost_giou: float = 1):
"""Creates the matcher
Params:
cost_class: This is the relative weight of the classification error in the matching cost
cost_bbox: This is the relative weight of the L1 error of the bounding box coordinates in the matching cost
cost_giou: This is the relative weight of the giou loss of the bounding box in the matching cost
"""
super().__init__()
self.cost_class = cost_class
self.cost_bbox = cost_bbox
self.cost_giou = cost_giou
assert cost_class != 0 or cost_bbox != 0 or cost_giou != 0, "all costs cant be 0"
@torch.no_grad() # 取消了梯度的产生，不参与回归
def forward(self, outputs, targets):
""" Performs the matching
Params:
outputs: This is a dict that contains at least these entries:
"pred_logits": Tensor of dim [batch_size, num_queries, num_classes] with the classification logits
"pred_boxes": Tensor of dim [batch_size, num_queries, 4] with the predicted box coordinates
targets: This is a list of targets (len(targets) = batch_size), where each target is a dict containing:
"labels": Tensor of dim [num_target_boxes] (where num_target_boxes is the number of ground-truth
objects in the target) containing the class labels
"boxes": Tensor of dim [num_target_boxes, 4] containing the target box coordinates
Returns:
A list of size batch_size, containing tuples of (index_i, index_j) where:
- index_i is the indices of the selected predictions (in order)
- index_j is the indices of the corresponding selected targets (in order)
For each batch element, it holds:
len(index_i) = len(index_j) = min(num_queries, num_target_boxes)
"""
bs, num_queries = outputs["pred_logits"].shape[:2]
# We flatten to compute the cost matrices in a batch
out_prob = outputs["pred_logits"].flatten(0, 1).softmax(-1) # [batch_size * num_queries, num_classes]
out_bbox = outputs["pred_boxes"].flatten(0, 1) # [batch_size * num_queries, 4]
# Also concat the target labels and boxes
tgt_ids = torch.cat([v["labels"] for v in targets])
tgt_bbox = torch.cat([v["boxes"] for v in targets])
print('0:', out_prob.shape)
print('1:', out_bbox.shape)
print('2:', tgt_ids.shape)
print('3:', tgt_bbox.shape)
# Compute the classification cost. Contrary to the loss, we don't use the NLL,
# but approximate it in 1 - proba[target class].
# The 1 is a constant that doesn't change the matching, it can be ommitted.
cost_class = -out_prob[:, tgt_ids] # 1 - proba[target class]. [batch_size * num_queries, target class]
print('4:', cost_class.shape)
# print('\n')
# Compute the L1 cost between boxes,求解正则项，L1范式
# 这个方法会对每个预测框与GT都进行误差计算。例如预测框N个，GT框M个。结果会有N*M个值（一个batch）
cost_bbox = torch.cdist(out_bbox, tgt_bbox, p=1)
print('5:', cost_bbox.shape)
# Compute the giou cost betwen boxes，省略常熟 1-generalized_box_iou
cost_giou = -generalized_box_iou(box_cxcywh_to_xyxy(out_bbox), box_cxcywh_to_xyxy(tgt_bbox))
print('6:', cost_giou.shape)
# Final cost matrix
C = self.cost_bbox * cost_bbox + self.cost_class * cost_class + self.cost_giou * cost_giou
print('7:', C.shape)
C = C.view(bs, num_queries, -1).cpu()
print('8:', C.shape)
sizes = [len(v["boxes"]) for v in targets] # 当前batch每张图像的目标GT数量，用于切分给每个图
print('9:', sizes)
# print('10:', C.split(sizes, -1))
indices = [linear_sum_assignment(c[i]) for i, c in enumerate(C.split(sizes, -1))]
for i, c in enumerate(C.split(sizes, -1)):
import numpy as np
cost_matrix = np.asarray(c[i])
# print('cost_matrix:', cost_matrix)
row_ind, col_ind = linear_sum_assignment(c[i])
for (row, col) in zip(row_ind, col_ind):
print(row, col, '***', cost_matrix[row][col])
print('11:', indices)
return [(torch.as_tensor(i, dtype=torch.int64), torch.as_tensor(j, dtype=torch.int64)) for i, j in indices]
# 匈牙利最优匹配，返回匹配索引
def build_matcher(args):
return HungarianMatcher(cost_class=args.set_cost_class, cost_bbox=args.set_cost_bbox, cost_giou=args.set_cost_giou)
把各阶段显示结果打印出来，查看如下：
0: torch.Size([100, 83])
1: torch.Size([100, 4])
2: torch.Size([3])
3: torch.Size([3, 4])
4: torch.Size([100, 3])
5: torch.Size([100, 3])
6: torch.Size([100, 3])
7: torch.Size([100, 3])
8: torch.Size([1, 100, 3])
9: [3]
10: (tensor([[[ 8.2781e+00, 5.7523e+00, 6.2704e+00],
[ 4.9576e+00, 6.8206e+00, 5.8613e+00],
[ 4.3011e+00, 1.1709e+00, 7.6130e+00],
[ 6.8157e+00, 1.0191e+01, 7.7051e+00],
[ 5.1331e+00, 2.5986e-01, 9.5993e+00],
[ 3.6728e+00, -3.1526e-01, 8.6469e+00],
[ 3.3763e+00, 6.1920e-01, 8.0840e+00],
[ 4.2639e+00, 3.7820e-01, 8.3531e+00],
[ 4.9786e+00, 1.4702e+00, 9.3967e+00],
[ 4.7091e+00, 1.9719e+00, 9.3294e+00],
[-1.9768e-01, 5.2733e+00, 4.0208e+00],
[ 7.9992e+00, 5.5239e+00, 6.2364e+00],
[-6.9660e-02, 4.1440e+00, 4.9423e+00],
[ 3.6899e+00, 8.7728e+00, 2.2498e-01],
[ 8.3986e-01, 5.0391e+00, 4.6867e+00],
[ 5.2078e+00, -1.0942e+00, 9.9988e+00],
[ 9.2502e+00, 6.9210e+00, 9.6557e+00],
[ 6.6826e+00, 1.0023e+01, 7.5727e+00],
[-4.2688e-01, 5.6021e+00, 3.7218e+00],
[ 1.5247e+00, 4.2876e+00, 6.6519e+00],
[-4.5765e-01, 5.3721e+00, 4.0550e+00],
[-1.0363e-01, 5.3039e+00, 3.8190e+00],
[ 7.0395e-01, 5.4264e+00, 3.7965e+00],
[ 4.5934e+00, 9.4687e+00, -9.1822e-01],
[ 6.6833e+00, 1.0135e+01, 7.5300e+00],
[-4.8056e-01, 5.0123e+00, 4.2394e+00],
[ 2.6458e-01, 5.1980e+00, 4.0117e+00],
[ 4.1776e+00, -2.8836e-01, 8.6328e+00],
[ 6.6063e+00, 9.5719e+00, 7.3849e+00],
[ 6.2134e+00, 3.5192e+00, 7.1582e+00],
[ 7.1327e+00, 4.5764e+00, 6.3685e+00],
[ 3.7398e+00, -9.6579e-02, 8.5241e+00],
[ 6.8127e+00, 4.1113e+00, 6.5262e+00],
[ 3.5170e+00, 8.8692e+00, 3.4571e+00],
[ 1.5750e-02, 4.7749e+00, 4.2767e+00],
[ 3.8504e+00, 6.3830e-01, 8.3046e+00],
[ 3.2140e+00, 8.3388e+00, 6.9667e-01],
[ 4.1912e+00, 8.4101e-01, 8.4362e+00],
[ 3.5930e+00, 1.1387e-01, 8.3090e+00],
[ 3.6005e+00, 8.4123e+00, 4.9653e+00],
[ 3.3979e+00, 8.6321e+00, 3.6201e+00],
[ 8.2190e-01, 3.8224e+00, 5.8881e+00],
[ 1.0869e+00, 3.9131e+00, 6.9495e+00],
[ 4.0383e+00, -1.6680e-01, 8.6249e+00],
[ 4.5527e+00, 1.3161e+00, 8.0127e+00],
[ 4.8717e+00, -1.1607e+00, 9.6060e+00],
[-2.3139e-01, 5.9623e+00, 3.7352e+00],
[-5.2189e-01, 5.1185e+00, 4.1926e+00],
[ 5.1255e+00, -1.0225e+00, 9.5479e+00],
[ 4.0998e+00, 8.8282e-01, 8.1521e+00],
[ 7.9094e+00, 5.9065e+00, 9.1802e+00],
[ 5.6728e-01, 4.0475e+00, 6.3995e+00],
[ 5.6560e+00, 2.6218e+00, 6.4698e+00],
[ 4.0612e+00, 1.8517e-01, 8.8067e+00],
[ 8.0742e+00, 5.5891e+00, 6.2239e+00],
[ 4.0400e+00, -4.1363e-01, 8.6222e+00],
[ 6.4105e+00, 3.6744e+00, 6.7804e+00],
[-1.7385e+00, 4.9004e+00, 4.6942e+00],
[ 3.7932e+00, -3.4077e-01, 8.6290e+00],
[ 1.5052e+00, 4.1979e+00, 6.7877e+00],
[ 4.9204e+00, 1.4836e-01, 9.7124e+00],
[ 4.0792e+00, -8.9709e-01, 8.9150e+00],
[ 1.2695e+00, 6.2007e+00, 6.3281e+00],
[ 4.7756e+00, -2.6407e-02, 9.4340e+00],
[ 6.8410e+00, 1.0208e+01, 7.6384e+00],
[ 3.7301e+00, 4.3986e-01, 8.0444e+00],
[ 4.9536e+00, 5.4823e-01, 9.4097e+00],
[ 4.6248e+00, -1.5043e+00, 9.4438e+00],
[ 4.8150e+00, 1.5749e+00, 8.2575e+00],
[ 2.6885e+00, 7.9073e+00, 1.2976e+00],
[ 4.5642e+00, 9.2036e+00, 5.3550e+00],
[ 5.4012e+00, -5.6389e-01, 1.0175e+01],
[ 4.2466e-01, 3.5950e+00, 5.6307e+00],
[ 4.5491e-02, 5.9585e+00, 3.3820e+00],
[-5.2628e-01, 5.4329e+00, 4.0332e+00],
[-4.2191e-01, 5.0971e+00, 4.0838e+00],
[ 4.4857e+00, 5.5204e-01, 8.3841e+00],
[ 8.6591e+00, 6.1527e+00, 6.5773e+00],
[ 4.1859e+00, 1.0231e+00, 8.2301e+00],
[ 4.3173e+00, 1.5263e+00, 8.0855e+00],
[ 1.9756e+00, 7.3597e+00, 1.8068e+00],
[ 3.7020e+00, -1.5050e-03, 8.4148e+00],
[ 1.2558e+00, 6.2132e+00, 2.9544e+00],
[ 5.5785e+00, 8.5805e+00, 6.9275e+00],
[ 4.4306e+00, 1.2772e+00, 7.9827e+00],
[ 6.9211e-01, 6.3655e+00, 2.7999e+00],
[ 4.4060e+00, 9.3261e+00, -1.0840e+00],
[-3.4285e-01, 3.9853e+00, 5.7059e+00],
[ 4.3454e+00, 4.4445e+00, 9.2292e+00],
[-1.5630e+00, 4.9668e+00, 4.5768e+00],
[-1.2383e+00, 4.4616e+00, 4.9227e+00],
[ 1.6393e-01, 4.6841e+00, 4.3523e+00],
[-5.9907e-01, 5.0048e+00, 4.3742e+00],
[ 5.1124e+00, 6.9615e+00, 6.0210e+00],
[ 3.0988e-03, 6.2910e+00, 3.5968e+00],
[-4.8344e-01, 4.9876e+00, 4.4395e+00],
[ 5.1149e+00, 1.9913e+00, 7.9950e+00],
[ 5.6246e+00, 8.8102e+00, 7.0050e+00],
[ 2.6130e-01, 3.6337e+00, 5.7934e+00],
[ 5.5709e+00, 8.9151e+00, 6.7141e+00]]]),)
cost_matrix: [[ 8.2780924e+00 5.7523308e+00 6.2704163e+00]
[ 4.9575653e+00 6.8206110e+00 5.8613491e+00]
[ 4.3011103e+00 1.1708755e+00 7.6129541e+00]
[ 6.8156848e+00 1.0191437e+01 7.7051115e+00]
[ 5.1331429e+00 2.5985980e-01 9.5993195e+00]
[ 3.6727586e+00 -3.1525683e-01 8.6469116e+00]
[ 3.3762672e+00 6.1919785e-01 8.0840225e+00]
[ 4.2639432e+00 3.7819684e-01 8.3531294e+00]
[ 4.9785528e+00 1.4702234e+00 9.3966799e+00]
[ 4.7090969e+00 1.9718820e+00 9.3293982e+00]
[-1.9767678e-01 5.2733393e+00 4.0208311e+00]
[ 7.9992228e+00 5.5239344e+00 6.2364326e+00]
[-6.9660187e-02 4.1440396e+00 4.9423199e+00]
[ 3.6898577e+00 8.7727966e+00 2.2497982e-01]
[ 8.3986056e-01 5.0390706e+00 4.6867170e+00]
[ 5.2078104e+00 -1.0942475e+00 9.9988441e+00]
[ 9.2502470e+00 6.9209762e+00 9.6557140e+00]
[ 6.6826372e+00 1.0023033e+01 7.5726542e+00]
[-4.2687911e-01 5.6021390e+00 3.7218261e+00]
[ 1.5246817e+00 4.2875733e+00 6.6519294e+00]
[-4.5764512e-01 5.3720751e+00 4.0549679e+00]
[-1.0362893e-01 5.3039198e+00 3.8189650e+00]
[ 7.0395356e-01 5.4264174e+00 3.7965493e+00]
[ 4.5933971e+00 9.4687033e+00 -9.1822195e-01]
[ 6.6832952e+00 1.0134920e+01 7.5299850e+00]
[-4.8055774e-01 5.0122609e+00 4.2393503e+00]
[ 2.6457596e-01 5.1979795e+00 4.0117393e+00]
[ 4.1775527e+00 -2.8836286e-01 8.6328087e+00]
[ 6.6063323e+00 9.5719442e+00 7.3848906e+00]
[ 6.2134027e+00 3.5191741e+00 7.1582074e+00]
[ 7.1326599e+00 4.5764084e+00 6.3685322e+00]
[ 3.7398467e+00 -9.6579432e-02 8.5240564e+00]
[ 6.8126822e+00 4.1112809e+00 6.5261931e+00]
[ 3.5170491e+00 8.8692226e+00 3.4571378e+00]
[ 1.5750408e-02 4.7749467e+00 4.2767425e+00]
[ 3.8504438e+00 6.3830268e-01 8.3045692e+00]
[ 3.2139552e+00 8.3388023e+00 6.9667470e-01]
[ 4.1911783e+00 8.4101266e-01 8.4362345e+00]
[ 3.5929587e+00 1.1386955e-01 8.3090458e+00]
[ 3.6005147e+00 8.4122725e+00 4.9652863e+00]
[ 3.3979454e+00 8.6320591e+00 3.6200943e+00]
[ 8.2189524e-01 3.8224022e+00 5.8880587e+00]
[ 1.0868909e+00 3.9131203e+00 6.9494729e+00]
[ 4.0383496e+00 -1.6679776e-01 8.6248922e+00]
[ 4.5527411e+00 1.3160551e+00 8.0127211e+00]
[ 4.8717475e+00 -1.1606574e+00 9.6059589e+00]
[-2.3138726e-01 5.9623160e+00 3.7351766e+00]
[-5.2189153e-01 5.1185379e+00 4.1925898e+00]
[ 5.1254988e+00 -1.0224745e+00 9.5478582e+00]
[ 4.0997610e+00 8.8282138e-01 8.1520863e+00]
[ 7.9093728e+00 5.9065194e+00 9.1801624e+00]
[ 5.6727898e-01 4.0474801e+00 6.3994970e+00]
[ 5.6560211e+00 2.6217759e+00 6.4698277e+00]
[ 4.0612087e+00 1.8516517e-01 8.8067341e+00]
[ 8.0741739e+00 5.5890713e+00 6.2239499e+00]
[ 4.0400124e+00 -4.1362917e-01 8.6222038e+00]
[ 6.4105182e+00 3.6743789e+00 6.7803993e+00]
[-1.7385412e+00 4.9003649e+00 4.6941576e+00]
[ 3.7932222e+00 -3.4077275e-01 8.6289921e+00]
[ 1.5052190e+00 4.1978951e+00 6.7876673e+00]
[ 4.9203916e+00 1.4835656e-01 9.7123594e+00]
[ 4.0792198e+00 -8.9708799e-01 8.9150419e+00]
[ 1.2695322e+00 6.2007360e+00 6.3281250e+00]
[ 4.7756453e+00 -2.6406884e-02 9.4340038e+00]
[ 6.8410158e+00 1.0207616e+01 7.6383772e+00]
[ 3.7301140e+00 4.3985701e-01 8.0444126e+00]
[ 4.9535618e+00 5.4822862e-01 9.4097462e+00]
[ 4.6247811e+00 -1.5043023e+00 9.4437799e+00]
[ 4.8150125e+00 1.5749331e+00 8.2575169e+00]
[ 2.6885109e+00 7.9073000e+00 1.2975867e+00]
[ 4.5641971e+00 9.2036409e+00 5.3550172e+00]
[ 5.4011717e+00 -5.6389362e-01 1.0175209e+01]
[ 4.2465532e-01 3.5949950e+00 5.6307044e+00]
[ 4.5490503e-02 5.9585238e+00 3.3820138e+00]
[-5.2628189e-01 5.4329481e+00 4.0332346e+00]
[-4.2191225e-01 5.0970836e+00 4.0838089e+00]
[ 4.4857378e+00 5.5203629e-01 8.3840704e+00]
[ 8.6590862e+00 6.1527267e+00 6.5773430e+00]
[ 4.1858683e+00 1.0230734e+00 8.2301025e+00]
[ 4.3172894e+00 1.5262530e+00 8.0854750e+00]
[ 1.9755765e+00 7.3597250e+00 1.8067718e+00]
[ 3.7019567e+00 -1.5050173e-03 8.4148455e+00]
[ 1.2558209e+00 6.2131724e+00 2.9543672e+00]
[ 5.5785089e+00 8.5805359e+00 6.9275179e+00]
[ 4.4305744e+00 1.2771857e+00 7.9826665e+00]
[ 6.9210845e-01 6.3654633e+00 2.7999129e+00]
[ 4.4059572e+00 9.3260632e+00 -1.0840294e+00]
[-3.4284663e-01 3.9852719e+00 5.7058821e+00]
[ 4.3453894e+00 4.4444699e+00 9.2292423e+00]
[-1.5629959e+00 4.9667954e+00 4.5767879e+00]
[-1.2383235e+00 4.4615502e+00 4.9226542e+00]
[ 1.6393489e-01 4.6840563e+00 4.3522577e+00]
[-5.9907275e-01 5.0048246e+00 4.3741856e+00]
[ 5.1123734e+00 6.9614697e+00 6.0210080e+00]
[ 3.0988455e-03 6.2910314e+00 3.5968068e+00]
[-4.8343736e-01 4.9875731e+00 4.4395008e+00]
[ 5.1148849e+00 1.9913349e+00 7.9949999e+00]
[ 5.6245937e+00 8.8101625e+00 7.0049658e+00]
[ 2.6129830e-01 3.6337457e+00 5.7934051e+00]
[ 5.5708518e+00 8.9150972e+00 6.7141314e+00]]
57 0 *** -1.7385412
67 1 *** -1.5043023
86 2 *** -1.0840294
11: [(array([57, 67, 86], dtype=int64), array([0, 1, 2]))]
上述直观到一张图看，如下这样：
其中核心部分，linear_sum_assignment的定义如下：
# Wrapper for the shortest augmenting path algorithm for solving the
# rectangular linear sum assignment problem. The original code was an
# implementation of the Hungarian algorithm (Kuhn-Munkres) taken from
# scikit-learn, based on original code by Brian Clapper and adapted to NumPy
# by Gael Varoquaux. Further improvements by Ben Root, Vlad Niculae, Lars
# Buitinck, and Peter Larsen.
#
# Copyright (c) 2008 Brian M. Clapper <bmc@clapper.org>, Gael Varoquaux
# Author: Brian M. Clapper, Gael Varoquaux
# License: 3-clause BSD
import numpy as np
from . import _lsap_module
def linear_sum_assignment(cost_matrix, maximize=False):
"""Solve the linear sum assignment problem.
The linear sum assignment problem is also known as minimum weight matching
in bipartite graphs. A problem instance is described by a matrix C, where
each C[i,j] is the cost of matching vertex i of the first partite set
(a "worker") and vertex j of the second set (a "job"). The goal is to find
a complete assignment of workers to jobs of minimal cost.
Formally, let X be a boolean matrix where :math:`X[i,j] = 1` iff row i is
assigned to column j. Then the optimal assignment has cost
.. math::
\\min \\sum_i \\sum_j C_{i,j} X_{i,j}
where, in the case where the matrix X is square, each row is assigned to
exactly one column, and each column to exactly one row.
This function can also solve a generalization of the classic assignment
problem where the cost matrix is rectangular. If it has more rows than
columns, then not every row needs to be assigned to a column, and vice
versa.
Parameters
----------
cost_matrix : array
The cost matrix of the bipartite graph.
maximize : bool (default: False)
Calculates a maximum weight matching if true.
Returns
-------
row_ind, col_ind : array
An array of row indices and one of corresponding column indices giving
the optimal assignment. The cost of the assignment can be computed
as ``cost_matrix[row_ind, col_ind].sum()``. The row indices will be
sorted; in the case of a square cost matrix they will be equal to
``numpy.arange(cost_matrix.shape[0])``.
Notes
-----
.. versionadded:: 0.17.0
References
----------
1. https://en.wikipedia.org/wiki/Assignment_problem
2. DF Crouse. On implementing 2D rectangular assignment algorithms.
*IEEE Transactions on Aerospace and Electronic Systems*,
52(4):1679-1696, August 2016, https://doi.org/10.1109/TAES.2016.140952
Examples
--------
>>> cost = np.array([[4, 1, 3], [2, 0, 5], [3, 2, 2]])
>>> from scipy.optimize import linear_sum_assignment
>>> row_ind, col_ind = linear_sum_assignment(cost)
>>> col_ind
array([1, 0, 2])
>>> cost[row_ind, col_ind].sum()
5
"""
cost_matrix = np.asarray(cost_matrix)
if len(cost_matrix.shape) != 2:
raise ValueError("expected a matrix (2-d array), got a %r array"
% (cost_matrix.shape,))
if not (np.issubdtype(cost_matrix.dtype, np.number) or
cost_matrix.dtype == np.dtype(np.bool)):
raise ValueError("expected a matrix containing numerical entries, got %s"
% (cost_matrix.dtype,))
if maximize:
cost_matrix = -cost_matrix
if np.any(np.isneginf(cost_matrix) | np.isnan(cost_matrix)):
raise ValueError("matrix contains invalid numeric entries")
cost_matrix = cost_matrix.astype(np.double)
a = np.arange(np.min(cost_matrix.shape))
# The algorithm expects more columns than rows in the cost matrix.
if cost_matrix.shape[1] < cost_matrix.shape[0]:
b = _lsap_module.calculate_assignment(cost_matrix.T)
indices = np.argsort(b)
return (b[indices], a[indices])
else:
b = _lsap_module.calculate_assignme看完这个，就觉得他好像就干了一件事：取最小的值，就是下面这段：
cost_matrix = cost_matrix.astype(np.double)
a = np.arange(np.min(cost_matrix.shape))用它给的这个栗子看看，输入代码是这样的：
import numpy as np
cost = np.array([[4, 1, 3],
[2, 0, 5],
[3, 2, 2]])
from scipy.optimize import linear_sum_assignment
row_ind, col_ind = linear_sum_assignment(cost)
print(row_ind, col_ind)
for (row, col) in zip(row_ind, col_ind):
print(row, col, '***', cost[row][col])打印的结果如下：（按行取最小值）
[0 1 2] [1 0 2]
0 1 *** 1
1 0 *** 2
2 2 *** 2三、总结
总结一下，匈牙利Hungarian算法在这里显得很蛮力。为了将预测的框，与标注的框找到1对1的最佳匹配，直接将预测的框M个，与标注的框N个，直接进行一一对应，组成一个M行N例的一个cost矩阵，其中每一个cost[m][n]就是一中组合对应形式。
最后，再按标注的N列，取出N个最优的预测值，这样构成预测与标注的一一对应，记为最佳匹配。
扫一扫
9万+
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
