直接放入nms yolov5源码,以一个二分类的模型举例说明,对nms的理解
def non_max_suppression(prediction, conf_thres=0.25, iou_thres=0.45, classes=None, agnostic=False, multi_label=False,
labels=(), max_det=100000,return_index = False):
"""Runs Non-Maximum Suppression (NMS) on inference results
Returns:
list of detections, on (n,6) tensor per image [xyxy, conf, cls]
"""
nc = prediction.shape[2] - 5 # number of classes
xc = prediction[..., 4] > conf_thres # candidates
# Checks
assert 0 <= conf_thres <= 1, f'Invalid Confidence threshold {conf_thres}, valid values are between 0.0 and 1.0'
assert 0 <= iou_thres <= 1, f'Invalid IoU {iou_thres}, valid values are between 0.0 and 1.0'
# Settings
min_wh, max_wh = 2, 4096 # (pixels) minimum and maximum box width and height
max_nms = 200000 # maximum number of boxes into torchvision.ops.nms()
time_limit = 60.0 # seconds to quit after
redundant = True # require redundant detections
multi_label &= nc > 1 # multiple labels per box (adds 0.5ms/img)
merge = False # use merge-NMS
t = time.time()
i = None
output = [torch.zeros((0, 6), device=prediction.device)] * prediction.shape[0]
for xi, x in enumerate(prediction): # image index, image inference
# Apply constraints
# x[((x[..., 2:4] < min_wh) | (x[..., 2:4] > max_wh)).any(1), 4] = 0 # width-height
x = x[xc[xi]] # confidence
# Cat apriori labels if autolabelling
if labels and len(labels[xi]):
l = labels[xi]
v = torch.zeros((len(l), nc + 5), device=x.device)
v[:, :4] = l[:, 1:5] # box
v[:, 4] = 1.0 # conf
v[range(len(l)), l[:, 0].long() + 5] = 1.0 # cls
x = torch.cat((x, v), 0)
# If none remain process next image
if not x.shape[0]:
continue
# Compute conf
x[:, 5:] *= x[:, 4:5] # conf = obj_conf * cls_conf
# Box (center x, center y, width, height) to (x1, y1, x2, y2)
box = xywh2xyxy(x[:, :4])
# Detections matrix nx6 (xyxy, conf, cls)
if multi_label:
i, j = (x[:, 5:] > conf_thres).nonzero(as_tuple=False).T
x = torch.cat((box[i], x[i, j + 5, None], j[:, None].float()), 1)
else: # best class only
conf, j = x[:, 5:].max(1, keepdim=True)
x = torch.cat((box, conf, j.float()), 1)[conf.view(-1) > conf_thres]
# Filter by class
if classes is not None:
x = x[(x[:, 5:6] == torch.tensor(classes, device=x.device)).any(1)]
# Apply finite constraint
# if not torch.isfinite(x).all():
# x = x[torch.isfinite(x).all(1)]
# Check shape
n = x.shape[0] # number of boxes
if not n: # no boxes
continue
elif n > max_nms: # excess boxes
x = x[x[:, 4].argsort(descending=True)[:max_nms]] # sort by confidence
# Batched NMS
c = x[:, 5:6] * (0 if agnostic else max_wh) # classes
boxes, scores = x[:, :4] + c, x[:, 4] # boxes (offset by class), scores
i = torchvision.ops.nms(boxes, scores, iou_thres) # NMS
if i.shape[0] > max_det: # limit detections
i = i[:max_det]
if merge and (1 < n < 3E3): # Merge NMS (boxes merged using weighted mean)
# update boxes as boxes(i,4) = weights(i,n) * boxes(n,4)
iou = box_iou(boxes[i], boxes) > iou_thres # iou matrix
weights = iou * scores[None] # box weights
x[i, :4] = torch.mm(weights, x[:, :4]).float() / weights.sum(1, keepdim=True) # merged boxes
if redundant:
i = i[iou.sum(1) > 1] # require redundancy
output[xi] = x[i]
if (time.time() - t) > time_limit:
print(f'WARNING: NMS time limit {time_limit}s exceeded')
break # time limit exceeded
if return_index == True:
return output, i
else:
return output
nc = prediction.shape[2] - 5 # number of classes
xc = prediction[..., 4] > conf_thres # candidates
prediction是网络模型的直接输出
输出其shape是(1, 50000, 7), 1表示的是图片的个数,50000表示是网络预测的候选框的个数,7表示一组数其意义如下:
由这个可以看出来,nc可以得到网络预测的类别个数。
xc = prediction[…, 4] > conf_thres # candidates
prediction[…, 4]是shape 为torch.Size([1, 50000]) tensor,prediction[…, 4]意义是取所有预测值的第5个值,表示目标框含有目标的概率值,整个表达式prediction[…, 4] > conf_thres,返回的是一个与prediction[…, 4]具有相同shape(1, 50000)的tensor,其每个值是True或者False,然后把这个tensor赋值给xc,所以xc的shape是(1, 50000),每个值是True或者False,用其值表达每个box的置信度是否大于或者小于conf_thres值。
输入参数conf_thres和iou_thres的check
assert 0 <= conf_thres <= 1, f'Invalid Confidence threshold {conf_thres}, valid values are between 0.0 and 1.0'
assert 0 <= iou_thres <= 1, f'Invalid IoU {iou_thres}, valid values are between 0.0 and 1.0'
这两行很简单,是check conf和iou是否在0和1之间,属于对参数的check。
min_wh, max_wh = 2, 4096 # (pixels) minimum and maximum box width and height
(像素值)最小和最大的box的宽和高
max_nms = 200000 # maximum number of boxes into torchvision.ops.nms()
输送到torchvision.ops.nms()接口中最大box总数量
time_limit = 60.0 # seconds to quit after
nms函数执行超时设置
redundant = True # require redundant detections
需要额外的检测
multi_label &= nc > 1 # multiple labels per box (adds 0.5ms/img)
merge = False # use merge-NMS
output = [torch.zeros((0, 6), device=prediction.device)] * prediction.shape[0]
output被定义成一个list,list的长度等于预测图片的个数prediction.shape[0],也就是1,每个list的元素是一个包含6个字段的tensor。
for xi, x in enumerate(prediction): # image index, image inference
xi是图片的index,其值是0,表示第0张图片,x是图片的推理结果,其shape是torch.Size([50000, 7]),这里巧妙的利用了enumerate将图片的index和图片的推理结果分离开,分别存放在了xi和x里面。本来prediction是一个shape为torch.Size([1, 50000, 7])的三维张量,张量的轴1是图片的index,张量的另外两个轴对应图片的推理结果,这里利用了enumerate将prediction的这两部分给分离开了。这行代码很巧妙。
x = x[xc[xi]] # confidence
xc是之前计算出的shape为(1, 50000)的tensor,每个值是True或者False,用其值表达每个box的置信度是否大于或者小于conf_thres值。xc[xi],[xi]这个是对xc tensor的取值方式,xi的值是0,则表示取第0张图片的所有True或者False值。xc[xi]的shape是torch.Size([50000]),与x的第0个维度一致,x[xc[xi]]也是对x tensor的取值方式,x[xc[xi]]整个表达式则表示对x第0个维度上xc[xi]所有为True的保留,为False的则舍弃,也就是对第i(这里i是0)张图片所有confidence值大于conf_thres的box取出,然后重新赋值给x。
# Cat apriori labels if autolabelling
if labels and len(labels[xi]):
l = labels[xi]
v = torch.zeros((len(l), nc + 5), device=x.device)
v[:, :4] = l[:, 1:5] # box
v[:, 4] = 1.0 # conf
v[range(len(l)), l[:, 0].long() + 5] = 1.0 # cls
x = torch.cat((x, v), 0)
这段代码不确定意思,暂时忽略,后面弄明白再补充。
# If none remain process next image
if not x.shape[0]:
continue
这里是检查通过conf_thres条件过滤后是否x.shape[0]大于0,如果有存在大于conf_thres的box,那么进一步处理。如果没有大于conf_thres的box,那么就处理下一张图片。
# Compute conf
x[:, 5:] *= x[:, 4:5] # conf = obj_conf * cls_conf
这行是计算每个框所预测的类别的概率,yolo论文的公式是
Pr ( Class i ∣ Object ) ∗ Pr ( Object ) ∗ I O U pred truth = Pr ( Class i ) ∗ I O U pred truth \operatorname{Pr}\left(\text { Class }_{i} \mid \text { Object }\right) * \operatorname{Pr}(\text { Object }) * \mathrm{IOU}_{\text {pred }}^{\text {truth }}=\operatorname{Pr}\left(\text { Class }_{i}\right) * \mathrm{IOU}_{\text {pred }}^{\text {truth }} Pr( Class i∣ Object )∗Pr( Object )∗IOUpred truth =Pr( Class i)∗IOUpred truth
这个引自于yolo论文gives us class-specific confidence scores for each
box. These scores encode both the probability of that class
appearing in the box and how well the predicted box fits the
object. 从这句话可以看出来,这个公式计算得到的scores值将每个类别的概率值以及预测框对目标框定范围的准确程度,而代码没有使用IOU这一部分,仅仅只做了类别的概率计算。x[:, 5:]对应的是 P r ( Class i ∣ Object ) {Pr}\left(\text { Class }_{i} \mid \text { Object }\right) Pr( Class i∣ Object ),x[:, 4:5]对应的是 P r ( Object ) {Pr}(\text { Object }) Pr( Object )。
而概率的计算也很巧妙,x[:, 5:]对应5000个2列tensor,x[:, 4:5],对应5000个1列tensor,x[:, 5:] *= x[:, 4:5]则表示为2列的每个元素和一列的每个元素分别相乘得到2列的元素, 元素再赋值给x[:, 5:]的两列。经过这一步计算x的后面两列(第6列,第7列)的值就是代表了目标的所在类别的confidence值了。
# Box (center x, center y, width, height) to (x1, y1, x2, y2)
box = xywh2xyxy(x[:, :4])
原来yolo的输出box的格式是center x, center y, width, height,经过这一步后将box的四个值表达为(x1, y1, x2, y2)
# Detections matrix nx6 (xyxy, conf, cls)
if multi_label:
i, j = (x[:, 5:] > conf_thres).nonzero(as_tuple=False).T
x = torch.cat((box[i], x[i, j + 5, None], j[:, None].float()), 1)
else: # best class only
conf, j = x[:, 5:].max(1, keepdim=True)
x = torch.cat((box, conf, j.float()), 1)[conf.view(-1) > conf_thres]
以上代码构建nx6的检测矩阵,对于multi_label,还没有明白,等明白了再补充。
conf, j = x[:, 5:].max(1, keepdim=True)
x = torch.cat((box, conf, j.float()), 1)[conf.view(-1) > conf_thres]
这里使用了tensor的max函数,这个函数的详细用法可以参考https://www.jianshu.com/p/3ed11362b54f,max的输入参数是1表示对每行求最大值,函数会返回两个tensor,第一个tensor是每行的最大值;第二个tensor是每行最大值的索引。keepdim可以参考https://blog.csdn.net/zylooooooooong/article/details/112576268,表示输出维度和输入维度是否一致,True则表示一致。
函数返回的conf的shape是torch.Size([50000, 1]),每一行是x[:, 5:]的最大值,j是每一行最大值的索引,这个索引最终用来表示成nms输出的每个类别的id。
x = torch.cat((box, conf, j.float()), 1)[conf.view(-1) > conf_thres]
这一行将box,conf,以及j 按照列cat成一个tensor作为网络的输出。[conf.view(-1) > conf_thres]则是筛选出出confidence值大于conf_thres所有box。
代码执行到这一行,基本就算是把置信度大于conf_thres的所有box给筛选出来了,筛选结果存放在x tensor里面。x的0到3列存放box,4列存放conf,5列存放class种类id。
# Filter by class
if classes is not None:
x = x[(x[:, 5:6] == torch.tensor(classes, device=x.device)).any(1)]
这一行是利用class进行过滤,筛选出指定的class,nms仅仅对指定的class进行nms。
# Check shape
n = x.shape[0] # number of boxes
if not n: # no boxes
continue
elif n > max_nms: # excess boxes
x = x[x[:, 4].argsort(descending=True)[:max_nms]] # sort by confidence
判断x中是否含有box,如果没有box,则进行下一张图片的 nms,如果box的个数超过最大nms个数则按照confidence值降序排列,取出最大nms个数的box做nms。
# Batched NMS
c = x[:, 5:6] * (0 if agnostic else max_wh) # classes
boxes, scores = x[:, :4] + c, x[:, 4] # boxes (offset by class), scores
i = torchvision.ops.nms(boxes, scores, iou_thres) # NMS
c = x[:, 5:6] * (0 if agnostic else max_wh)这行代码是多类别中应用NMS具体意义可以参考https://blog.csdn.net/flyfish1986/article/details/119177472。
多类别NMS(非极大值抑制)的处理策略是为了让每个类都能独立执行NMS,在所有的边框上添加一个偏移量。偏移量仅取决于类的ID(也就是x[:, 5:6]),并且足够大,以便来自不同类的框不会重叠。
boxes, scores = x[:, :4] + c, x[:, 4] # boxes (offset by class), scores
这行取出boxes和scores,boxes添加了偏移量c,不通过类别的偏移量大小不一致。
i = torchvision.ops.nms(boxes, scores, iou_thres) # NMS
调用torch自带的nms接口实现重叠框的抑制,函数返回的是一个tensor i
Tensor: int64 tensor with the indices of the elements that have been kept
by NMS, sorted in decreasing order of scores
i的意义是,整型64张量,指示被保留的框的index,另外是按照得分(置信度)从高到低排列。
if i.shape[0] > max_det: # limit detections
i = i[:max_det]
判断是否超过最大nms检测个数,如果超过,则去掉置信度低的。
if merge and (1 < n < 3E3): # Merge NMS (boxes merged using weighted mean)
# update boxes as boxes(i,4) = weights(i,n) * boxes(n,4)
iou = box_iou(boxes[i], boxes) > iou_thres # iou matrix
weights = iou * scores[None] # box weights
x[i, :4] = torch.mm(weights, x[:, :4]).float() / weights.sum(1, keepdim=True) # merged boxes
if redundant:
i = i[iou.sum(1) > 1] # require redundancy
这部分暂时没懂,后继懂了再更新
output[xi] = x[i]
x[i]是利用nms的结果i,取出所有nms 结果i对应的box,然后将结果保存到xi张图片对应的output里面。
if (time.time() - t) > time_limit:
print(f'WARNING: NMS time limit {time_limit}s exceeded')
break # time limit exceeded
这行是判断nms是否超时,如果超时,则直接跳出for循环,不进行下一张图片的nms
return output
返回结果,结果是存放在output里面了,output 对应的是一个list,list一个元素对应一张图片的nms结果,对于本文举例的二分类,只有一张图片,所以可以查看output[0],确认nms的输出结果,其shape是torch.Size([1892, 6]),如果打印其中一行,则对应的格式如下,也就是nms结果,每一行的意义: