CascadeRPN代码略读

目录

训练代码

train

parse_args

main

train

train_detector

batch_processor/parse_losses

build_optimizer

_non_dist_train/_dist_train

数据

CocoDataset

load_annotations

getitem

_filter_imgs

get_ann_info

_parse_ann_info

CustomDataset

getitem

prepare_train_img

模型

Cascade RPN

架构

forward_train

用到了rpn_head的方法

CascadeRPNHead

AdaptiveConv

forward_single

loss1

get_bboxes_single

CascadeAnchorHead

forward

init_anchors

refine_bboxes

loss

BaseDetector

forward(self, img, img_meta, return_loss=True, **kwargs):

RPNTestMixin

simple_test_rpn

aug_test_rpn

附录代码

RegionAssigner代码

RandomSampler代码

MaxIoUAssigner代码

训练代码

train

tools，训练代码

parse_args

用来接收各种参数

main

建立模型

model = build_detector(cfg.model, train_cfg=cfg.train_cfg, test_cfg=cfg.test_cfg)

建立数据集

datasets = [build_dataset(cfg.data.train)]

调用训练api

train_detector(model,datasets,cfg,distributed=distributed,validate=args.validate,logger=logger)

train

mmdet/api，训练代码接口

train_detector

调用分布训练_dist_train或者非分布训练_non_dist_train

batch_processor/parse_losses

整合batch中的loss

build_optimizer

配置optimizer

_non_dist_train/_dist_train

训练 主要看一下 _non_dist_train

载入数据

datasets = [build_dataset(cfg.data.train)]

配置dataloader,采样是GroupSampler，然后是官方的dataloader

data_loaders = [build_dataloader(ds, cfg.data.imgs_per_gpu, cfg.data.workers_per_gpu, dist=True)for ds in dataset]

建立model

model = MMDistributedDataParallel(model.cuda())

建立optimizer，函数在本文件里就有

optimizer = build_optimizer(model, cfg.optimizer)

配置runner

runner = Runner(model, batch_processor, optimizer, cfg.work_dir,cfg.log_level)

Runner是mmcv自带

class Runner(object):"""A training helper for PyTorch.Args:model (:obj:`torch.nn.Module`): The model to be run.batch_processor (callable): A callable method that process a databatch. The interface of this method should be`batch_processor(model, data, train_mode) -> dict`optimizer (dict or :obj:`torch.optim.Optimizer`): If it is a dict,runner will construct an optimizer according to it.work_dir (str, optional): The working directory to save checkpointsand logs.log_level (int): Logging level.logger (:obj:`logging.Logger`): Custom logger. If `None`, use thedefault logger.meta (dict | None): A dict records some import information such asenvironment info and seed, which will be logged in logger hook."""

gogogo 开始跑

runner.run(data_loaders, cfg.workflow, cfg.total_epochs)

数据

DATASET是Registry，保存了CocoDataset，PIPLINE在CustomDataset，这个类是数据类的父类，后面有

def build_dataset(cfg, default_args=None):...dataset = build_from_cfg(cfg, DATASETS, default_args)...return dataset

CocoDataset

继承了CustomDataset

load_annotations

读入图片id
读入类别并转换成label
返回图片信息

getitem

直接继承的父类

_filter_imgs

改写，加了一个过滤没有gt的功能

get_ann_info

读取img[index]的annotaion，送入_parse_ann_info处理

_parse_ann_info

这是处理数据的主要部分

gt_bboxes = []
gt_labels = []
gt_bboxes_ignore = []
gt_masks_ann = []...ann = dict(bboxes=gt_bboxes,labels=gt_labels,bboxes_ignore=gt_bboxes_ignore,masks=gt_masks_ann,seg_map=seg_map)

CustomDataset

getitem

继承DATASET 分别对test train数据进行处理

prepare_test_img/prepare_train_img

prepare_train_img

get_ann_info获得标注

pre_pipeline 给结果中加入 标注信息（图片地址）

pipeline PIPELIN处理

self.pipeline = Compose(pipeline)

模型

CascadeRPNHead -> (继承)-> CascadeAnchorHead -> (loss中的函数)-> region_anchor_target ->（region_anchor_target_single中类调用bbox_assigner.assign）-> RegionAssigner.assign ->返回一个AssignResult类

build detector

def build(cfg, registry, default_args=None):if isinstance(cfg, list):modules = [build_from_cfg(cfg_, registry, default_args) for cfg_ in cfg]return nn.Sequential(*modules)else:return build_from_cfg(cfg, registry, default_args)

Cascade RPN

继承了BaseDetector作为模板。

继承了RPNTestMixin，没看出来有啥用，此类源码看起来是用来跑RPN的test

架构

self.backbone = builder.build_backbone(backbone)
self.neck = builder.build_neck(neck) if neck is not None else None
self.rpn_head = nn.ModuleList()
for head in rpn_head:self.rpn_head.append(builder.build_head(head))

forward_train

def forward_train(self, img, img_meta, gt_bboxes=None):#FPN+RESNETx = self.extract_feat(img)featmap_sizes = [featmap.size()[-2:] for featmap in x]#generate anchoranchor_list, valid_flag_list = self.rpn_head[0].init_anchors(featmap_sizes, img_meta)losses = dict()for i in range(self.num_stages):# num_stage == 2rpn_train_cfg = self.train_cfg.rpn[i]rpn_head = self.rpn_head[i]if rpn_head.feat_adapt:#not in first stageoffset_list = anchor_offset(anchor_list,rpn_head.anchor_strides,featmap_sizes)else:offset_list = Nonex, cls_score, bbox_pred = rpn_head(x, offset_list)rpn_loss_inputs = (anchor_list, valid_flag_list, cls_score,bbox_pred, gt_bboxes, img_meta, rpn_train_cfg)stage_loss = rpn_head.loss(*rpn_loss_inputs)for name, value in stage_loss.items():losses['s{}.{}'.format(i, name)] = value# refine boxesif i < self.num_stages - 1:anchor_list = rpn_head.refine_bboxes(anchor_list, bbox_pred,img_meta)return losses

用到了rpn_head的方法

1. init_anchors
2. loss，loss的input：rpn_loss_inputs = (anchor_list,valid_flag_list, cls_score, bbox_pred, gt_bboxes, img_meta, rpn_train_cfg)
3. refine_bboxes
4. anchor_offset（这个不是rpn_head）anchor_offset(anchor_list, rpn_head.anchor_strides,featmap_sizes)

CascadeRPNHead

继承了CascadeAnchorHead作为模板

AdaptiveConv

adapt=True则为 DeformConv，否则为普通的dilation

forward_single

forward 继承父类 forward

def forward_single(self, x, offset):bridged_x = xx = self.relu(self.rpn_conv(x, offset))#AdaptiveConv#第一次传出过AdaptiveConv的feature 第二次传出不过的if self.bridged_feature:bridged_x = x  # update featurecls_score = self.rpn_cls(x) if self.with_cls else Nonebbox_pred = self.rpn_reg(x)return bridged_x, cls_score, bbox_pred

loss1

loss 继承父类 loss

加了一个是否分类的判断,第一层只回归，第二次分类+回归

if self.with_cls:return dict(loss_rpn_cls=losses['loss_cls'],loss_rpn_reg=losses['loss_reg'])
return dict(loss_rpn_reg=losses['loss_reg'])

get_bboxes_single

暂时涉及不到，这个是生成后续proposal的，和rpn无关

CascadeAnchorHead

forward

def forward(self, feats, offset_list=None):if offset_list is None:offset_list = [None for _ in range(len(feats))]return multi_apply(self.forward_single, feats, offset_list)

init_anchors

valid_flag_list list :判断anhcor在合法范围内

生成的 anchor_list, valid_flag_list大小为：
img*FPN*anchors

#anchor_list
#torch.Size([60800, 4])
#torch.Size([15200, 4])
#torch.Size([3800, 4])
#torch.Size([950, 4])      
#torch.Size([247, 4]) #valid_flag_list
#torch.Size([60800])
#torch.Size([15200])
#torch.Size([3800])
#torch.Size([950])      
#torch.Size([247]) 

测试了一下，估计是config文件里的pad设置，所有的valid_flag_list都是1

refine_bboxes

只在第一层起作用，利用第一层的regression微调anchor
delta2bbox计算就是正常的计算方式

\delta_x = (t_x-a_x)/a_w\delta_y = (t_y-a_y)/a_h\delta_w = log(t_w/a_w)\delta_h = log(t_h/a_h)

so δ去预测变化得到a' 尽可能地向t靠近

a^{'}_x = \delta_x*a_w+a_xa^{'}_y = \delta_y*a_h+a_ya^{'}_w = exp(\delta_w)*a_wa^{'}_h = exp(\delta_h)*a_h

loss

assigner 第一阶段是RegionAssigner，第二阶段是MaxIoUAssigner

RegionAssigner

这阶段不用找到负样本 只需要正样本来返回回归就可以了

def region_anchor_target(anchor_list,valid_flag_list,gt_bboxes_list,img_metas,featmap_sizes,anchor_scale,anchor_strides,target_means,target_stds,cfg,gt_bboxes_ignore_list=None,gt_labels_list=None,label_channels=1,sampling=True,unmap_outputs=True):# TODO add docs"""Compute regression and classification targets for anchors.Args:anchor_list (list[list]): Multi level anchors of each image.valid_flag_list (list[list]): Multi level valid flags of each image.gt_bboxes_list (list[Tensor]): Ground truth bboxes of each image.img_metas (list[dict]): Meta info of each image.target_means (Iterable): Mean value of regression targets.target_stds (Iterable): Std value of regression targets.cfg (dict): RPN train configs.Returns:tuple"""

region_anchor_target_single生成all_labels, all_label_weights, all_bbox_targets, all_bbox_weights,pos_inds_list, neg_inds_list，变换每张图的各个变量到每个level上

region_anchor_target_single

调用RegionAssigner

at/DATA7_DB7/data/tfhou/rpn/Cascade-RPN/mmdet/core/bbox/assigners/region_assigner.py

代码

代码里有文档，这里仅列出摘要

each bbox will be assigned with -1, 0, or a positive number. -1 means don't care, 0 means negative sample

1. Assign every anchor to 0 (negative)

For each gt_bboxes:

1. Compute ignore flags based on ignore_region then
   assign -1 to anchors w.r.t. ignore flags
2. Compute pos flags based on center_region then
   assign gt_bboxes to anchors w.r.t. pos flags
3. Compute ignore flags based on adjacent anchor lvl then
   assign -1 to anchors w.r.t. ignore flags
4. Assign anchor outside of image to -1

返回的result为num_gtsgt数量， gt_inds 0，-1，1标注
第一阶段sampling为False

第一阶段就是回归loss，并且对所有的anchor进行了微调，并且对特征进行adptive的卷积进入下一层

MaxIoUAssigner

第二阶段首先计算anchor_offset，根据现在的anchor位置和预定的位置进行计算

正常的iou寻找正负样本

def anchor_target(anchor_list,valid_flag_list,gt_bboxes_list,img_metas,target_means,target_stds,cfg,gt_bboxes_ignore_list=None,gt_labels_list=None,label_channels=1,sampling=True,unmap_outputs=True):"""Compute regression and classification targets for anchors.Args:anchor_list (list[list]): Multi level anchors of each image.valid_flag_list (list[list]): Multi level valid flags of each image.gt_bboxes_list (list[Tensor]): Ground truth bboxes of each image.img_metas (list[dict]): Meta info of each image.target_means (Iterable): Mean value of regression targets.target_stds (Iterable): Std value of regression targets.cfg (dict): RPN train configs.Returns:tuple"""

由于有了采样，进入assign_and_sample函数

bbox_assigner = build_assigner(cfg.assigner)
bbox_sampler = build_sampler(cfg.sampler)
assign_result = bbox_assigner.assign(bboxes, gt_bboxes,gt_bboxes_ignore,gt_labels)
sampling_result = bbox_sampler.sample(assign_result, bboxes,gt_bboxes,gt_labels)

assigner为MaxIoUAssigner

sampling为RandomSampler

BaseDetector

forward(self, img, img_meta, return_loss=True, **kwargs):

forward_train没写 留给儿砸

def forward(self, img, img_meta, return_loss=True, **kwargs):if return_loss:return self.forward_train(img, img_meta, **kwargs)else:return self.forward_test(img, img_meta, **kwargs)

RPNTestMixin

简单跑个rpn前向吧

simple_test_rpn

aug_test_rpn

附录代码

RegionAssigner代码

import torchfrom .assign_result import AssignResult
from .base_assigner import BaseAssignerdef calc_region(bbox, ratio, stride, featmap_size=None):# Base anchor locates in (stride - 1) * 0.5f_bbox = (bbox - (stride - 1) * 0.5) / stridex1 = torch.round((1 - ratio) * f_bbox[0] + ratio * f_bbox[2])y1 = torch.round((1 - ratio) * f_bbox[1] + ratio * f_bbox[3])x2 = torch.round(ratio * f_bbox[0] + (1 - ratio) * f_bbox[2])y2 = torch.round(ratio * f_bbox[1] + (1 - ratio) * f_bbox[3])if featmap_size is not None:x1 = x1.clamp(min=0, max=featmap_size[1] - 1)y1 = y1.clamp(min=0, max=featmap_size[0] - 1)x2 = x2.clamp(min=0, max=featmap_size[1] - 1)y2 = y2.clamp(min=0, max=featmap_size[0] - 1)return (x1, y1, x2, y2)def anchor_ctr_inside_region_flags(anchors, stride, region):x1, y1, x2, y2 = regionf_anchors = (anchors - (stride - 1) * 0.5) / stridex = (f_anchors[:, 0] + f_anchors[:, 2]) * 0.5y = (f_anchors[:, 1] + f_anchors[:, 3]) * 0.5flags = (x >= x1) & (x <= x2) & (y >= y1) & (y <= y2)return flagsdef anchor_outside_flags(flat_anchors,valid_flags,img_shape,allowed_border=0):img_h, img_w = img_shape[:2]if allowed_border >= 0:inside_flags = valid_flags & \(flat_anchors[:, 0] >= -allowed_border) & \(flat_anchors[:, 1] >= -allowed_border) & \(flat_anchors[:, 2] < img_w + allowed_border) & \(flat_anchors[:, 3] < img_h + allowed_border)else:inside_flags = valid_flagsoutside_flags = ~inside_flagsreturn outside_flagsclass RegionAssigner(BaseAssigner):"""Assign a corresponding gt bbox or background to each bbox.Each proposals will be assigned with `-1`, `0`, or a positive integerindicating the ground truth index.- -1: don't care- 0: negative sample, no assigned gt- positive integer: positive sample, index (1-based) of assigned gtArgs:pos_iou_thr (float): IoU threshold for positive bboxes.neg_iou_thr (float or tuple): IoU threshold for negative bboxes.min_pos_iou (float): Minimum iou for a bbox to be considered as apositive bbox. Positive samples can have smaller IoU thanpos_iou_thr due to the 4th step (assign max IoU sample to each gt)."""# TODO update docsdef __init__(self, center_ratio=0.2, ignore_ratio=0.5):self.center_ratio = center_ratioself.ignore_ratio = ignore_ratiodef assign(self,mlvl_anchors,mlvl_valid_flags,gt_bboxes,img_meta,featmap_sizes,anchor_scale,anchor_strides,gt_bboxes_ignore=None,gt_labels=None,allowed_border=0):"""Assign gt to anchors.This method assign a gt bbox to every bbox (proposal/anchor), each bboxwill be assigned with -1, 0, or a positive number. -1 means don't care,0 means negative sample, positive number is the index (1-based) ofassigned gt.The assignment is done in following steps, the order matters.1. Assign every anchor to 0 (negative)For each gt_bboxes:2. Compute ignore flags based on ignore_region thenassign -1 to anchors w.r.t. ignore flags3. Compute pos flags based on center_region thenassign gt_bboxes to anchors w.r.t. pos flags4. Compute ignore flags based on adjacent anchor lvl thenassign -1 to anchors w.r.t. ignore flags5. Assign anchor outside of image to -1Args:bboxes (Tensor): Bounding boxes to be assigned, shape(n, 4).gt_bboxes (Tensor): Groundtruth boxes, shape (k, 4).gt_bboxes_ignore (Tensor, optional): Ground truth bboxes that arelabelled as `ignored`, e.g., crowd boxes in COCO.gt_labels (Tensor, optional): Label of gt_bboxes, shape (k, ).Returns::obj:`AssignResult`: The assign result."""# TODO support gt_bboxes_ignoreif gt_bboxes_ignore is not None:raise NotImplementedErrorif gt_bboxes.shape[0] == 0:raise ValueError('No gt bboxes')num_gts = gt_bboxes.shape[0]num_lvls = len(mlvl_anchors)r1 = (1 - self.center_ratio) / 2  #0.4r2 = (1 - self.ignore_ratio) / 2  #0.25scale = torch.sqrt((gt_bboxes[:, 2] - gt_bboxes[:, 0] + 1) *(gt_bboxes[:, 3] - gt_bboxes[:, 1] + 1))min_anchor_size = scale.new_full((1, ), float(anchor_scale * anchor_strides[0]))#find the gt in which leveltarget_lvls = torch.floor(torch.log2(scale) - torch.log2(min_anchor_size) + 0.5)target_lvls = target_lvls.clamp(min=0, max=num_lvls - 1).long()# 1. assign 0 (negative) by defaultmlvl_assigned_gt_inds = []mlvl_ignore_flags = []for lvl in range(num_lvls):h, w = featmap_sizes[lvl]assert h * w == mlvl_anchors[lvl].shape[0]assigned_gt_inds = gt_bboxes.new_full((h * w, ),0,dtype=torch.long)ignore_flags = torch.zeros_like(assigned_gt_inds)mlvl_assigned_gt_inds.append(assigned_gt_inds)mlvl_ignore_flags.append(ignore_flags)for gt_id in range(num_gts):lvl = target_lvls[gt_id].item()featmap_size = featmap_sizes[lvl]stride = anchor_strides[lvl]anchors = mlvl_anchors[lvl]gt_bbox = gt_bboxes[gt_id, :4]# Compute regions# the different regions just depends on gt# func: calc_region# w & h become 0.5 for ignore / 0.2 for ctrignore_region = calc_region(gt_bbox, r2, stride, featmap_size)#print('ig',gt_bbox,ignore_region)ctr_region = calc_region(gt_bbox, r1, stride, featmap_size)#print('cr',gt_bbox,ctr_region)# 2. Assign -1 to ignore flagsignore_flags = anchor_ctr_inside_region_flags(anchors, stride, ignore_region)mlvl_assigned_gt_inds[lvl][ignore_flags > 0] = -1# 3. Assign gt_bboxes to pos flagspos_flags = anchor_ctr_inside_region_flags(anchors, stride,ctr_region)mlvl_assigned_gt_inds[lvl][pos_flags > 0] = gt_id + 1# 4. Assign -1 to ignore adjacent lvlif lvl > 0:d_lvl = lvl - 1d_anchors = mlvl_anchors[d_lvl]d_featmap_size = featmap_sizes[d_lvl]d_stride = anchor_strides[d_lvl]#print('!',d_stride,r2)# r2 & d_stride?d_ignore_region = calc_region(gt_bbox, d_stride, r2,d_featmap_size)#print('!',gt_bbox,d_ignore_region)# x1 > x2 ?# tensor([473.8398, 384.1748, 804.3822, 467.8830], device='cuda:0') # (tensor(167., dvice='cuda:0'), # tensor(99., device='cuda:0'), # tensor(0., device='cuda:0'), # tensor(0., device='cuda:0'))ignore_flags = anchor_ctr_inside_region_flags(d_anchors, d_stride, d_ignore_region)mlvl_ignore_flags[d_lvl][ignore_flags > 0] = 1if lvl < num_lvls - 1:u_lvl = lvl + 1u_anchors = mlvl_anchors[u_lvl]u_featmap_size = featmap_sizes[u_lvl]u_stride = anchor_strides[u_lvl]#print('?',u_stride,r2)# r2 & u_stride?u_ignore_region = calc_region(gt_bbox, u_stride,r2,u_featmap_size)#print('?',gt_bbox,u_ignore_region)ignore_flags = anchor_ctr_inside_region_flags(u_anchors, u_stride, u_ignore_region)mlvl_ignore_flags[u_lvl][ignore_flags > 0] = 1# 4. (cont.) Assign -1 to ignore adjacent lvlfor lvl in range(num_lvls):ignore_flags = mlvl_ignore_flags[lvl]mlvl_assigned_gt_inds[lvl][ignore_flags > 0] = -1# 5. Assign -1 to anchor outside of imageflat_assigned_gt_inds = torch.cat(mlvl_assigned_gt_inds)flat_anchors = torch.cat(mlvl_anchors)flat_valid_flags = torch.cat(mlvl_valid_flags)assert (flat_assigned_gt_inds.shape[0] == flat_anchors.shape[0] ==flat_valid_flags.shape[0])outside_flags = anchor_outside_flags(flat_anchors, flat_valid_flags,img_meta['img_shape'],allowed_border)flat_assigned_gt_inds[outside_flags] = -1if gt_labels is not None:assigned_labels = torch.zeros_like(flat_assigned_gt_inds)pos_flags = assigned_gt_inds > 0assigned_labels[pos_flags] = gt_labels[flat_assigned_gt_inds[pos_flags] - 1]else:assigned_labels = Nonereturn AssignResult(num_gts, flat_assigned_gt_inds, None, labels=assigned_labels)

RandomSampler代码

import numpy as np
import torchfrom .base_sampler import BaseSamplerclass RandomSampler(BaseSampler):#BaseSampler has function sampl;e and it use pos_sample and neg_sample # sample num is 256*0.5==128 but pos is always less than 128 neg is about 50000+def __init__(self,num,pos_fraction,neg_pos_ub=-1,add_gt_as_proposals=True,**kwargs):super(RandomSampler, self).__init__(num, pos_fraction, neg_pos_ub,add_gt_as_proposals)@staticmethoddef random_choice(gallery, num):"""Random select some elements from the gallery.It seems that Pytorch's implementation is slower than numpy so we usenumpy to randperm the indices."""assert len(gallery) >= numif isinstance(gallery, list):gallery = np.array(gallery)cands = np.arange(len(gallery))np.random.shuffle(cands)rand_inds = cands[:num]if not isinstance(gallery, np.ndarray):rand_inds = torch.from_numpy(rand_inds).long().to(gallery.device)return gallery[rand_inds]def _sample_pos(self, assign_result, num_expected, **kwargs):"""Randomly sample some positive samples."""pos_inds = torch.nonzero(assign_result.gt_inds > 0)if pos_inds.numel() != 0:pos_inds = pos_inds.squeeze(1)if pos_inds.numel() <= num_expected:return pos_indselse:return self.random_choice(pos_inds, num_expected)def _sample_neg(self, assign_result, num_expected, **kwargs):"""Randomly sample some negative samples."""neg_inds = torch.nonzero(assign_result.gt_inds == 0)if neg_inds.numel() != 0:neg_inds = neg_inds.squeeze(1)if len(neg_inds) <= num_expected:return neg_indselse:return self.random_choice(neg_inds, num_expected)

MaxIoUAssigner代码

import torchfrom ..geometry import bbox_overlaps
from .assign_result import AssignResult
from .base_assigner import BaseAssignerclass MaxIoUAssigner(BaseAssigner):"""Assign a corresponding gt bbox or background to each bbox.Each proposals will be assigned with `-1`, `0`, or a positive integerindicating the ground truth index.- -1: don't care- 0: negative sample, no assigned gt- positive integer: positive sample, index (1-based) of assigned gtArgs:pos_iou_thr (float): IoU threshold for positive bboxes.neg_iou_thr (float or tuple): IoU threshold for negative bboxes.min_pos_iou (float): Minimum iou for a bbox to be considered as apositive bbox. Positive samples can have smaller IoU thanpos_iou_thr due to the 4th step (assign max IoU sample to each gt).gt_max_assign_all (bool): Whether to assign all bboxes with the samehighest overlap with some gt to that gt.ignore_iof_thr (float): IoF threshold for ignoring bboxes (if`gt_bboxes_ignore` is specified). Negative values mean notignoring any bboxes.ignore_wrt_candidates (bool): Whether to compute the iof between`bboxes` and `gt_bboxes_ignore`, or the contrary."""def __init__(self,pos_iou_thr,neg_iou_thr,min_pos_iou=.0,gt_max_assign_all=True,ignore_iof_thr=-1,ignore_wrt_candidates=True):self.pos_iou_thr = pos_iou_thrself.neg_iou_thr = neg_iou_thrself.min_pos_iou = min_pos_iouself.gt_max_assign_all = gt_max_assign_allself.ignore_iof_thr = ignore_iof_thrself.ignore_wrt_candidates = ignore_wrt_candidatesdef assign(self, bboxes, gt_bboxes, gt_bboxes_ignore=None, gt_labels=None):"""Assign gt to bboxes.This method assign a gt bbox to every bbox (proposal/anchor), each bboxwill be assigned with -1, 0, or a positive number. -1 means don't care,0 means negative sample, positive number is the index (1-based) ofassigned gt.The assignment is done in following steps, the order matters.1. assign every bbox to -12. assign proposals whose iou with all gts < neg_iou_thr to 03. for each bbox, if the iou with its nearest gt >= pos_iou_thr,assign it to that bbox4. for each gt bbox, assign its nearest proposals (may be more thanone) to itselfArgs:bboxes (Tensor): Bounding boxes to be assigned, shape(n, 4).gt_bboxes (Tensor): Groundtruth boxes, shape (k, 4).gt_bboxes_ignore (Tensor, optional): Ground truth bboxes that arelabelled as `ignored`, e.g., crowd boxes in COCO.gt_labels (Tensor, optional): Label of gt_bboxes, shape (k, ).Returns::obj:`AssignResult`: The assign result."""if bboxes.shape[0] == 0 or gt_bboxes.shape[0] == 0:raise ValueError('No gt or bboxes')bboxes = bboxes[:, :4]# def bbox_overlaps(bboxes1, bboxes2, mode='iou', is_aligned=False)#calculate the iouoverlaps = bbox_overlaps(gt_bboxes, bboxes)if (self.ignore_iof_thr > 0) and (gt_bboxes_ignore is not None) and (gt_bboxes_ignore.numel() > 0):if self.ignore_wrt_candidates:ignore_overlaps = bbox_overlaps(bboxes, gt_bboxes_ignore, mode='iof')ignore_max_overlaps, _ = ignore_overlaps.max(dim=1)else:ignore_overlaps = bbox_overlaps(gt_bboxes_ignore, bboxes, mode='iof')ignore_max_overlaps, _ = ignore_overlaps.max(dim=0)overlaps[:, ignore_max_overlaps > self.ignore_iof_thr] = -1assign_result = self.assign_wrt_overlaps(overlaps, gt_labels)return assign_resultdef assign_wrt_overlaps(self, overlaps, gt_labels=None):"""Assign w.r.t. the overlaps of bboxes with gts.Args:overlaps (Tensor): Overlaps between k gt_bboxes and n bboxes,shape(k, n).gt_labels (Tensor, optional): Labels of k gt_bboxes, shape (k, ).Returns::obj:`AssignResult`: The assign result."""if overlaps.numel() == 0:raise ValueError('No gt or proposals')num_gts, num_bboxes = overlaps.size(0), overlaps.size(1)# 1. assign -1 by defaultassigned_gt_inds = overlaps.new_full((num_bboxes, ),-1,dtype=torch.long)# for each anchor, which gt best overlaps with it# for each anchor, the max iou of all gtsmax_overlaps, argmax_overlaps = overlaps.max(dim=0)# for each gt, which anchor best overlaps with it# for each gt, the max iou of all proposalsgt_max_overlaps, gt_argmax_overlaps = overlaps.max(dim=1)# 2. assign negative: belowif isinstance(self.neg_iou_thr, float):assigned_gt_inds[(max_overlaps >= 0)& (max_overlaps < self.neg_iou_thr)] = 0elif isinstance(self.neg_iou_thr, tuple):assert len(self.neg_iou_thr) == 2assigned_gt_inds[(max_overlaps >= self.neg_iou_thr[0])& (max_overlaps < self.neg_iou_thr[1])] = 0# 3. assign positive: above positive IoU thresholdpos_inds = max_overlaps >= self.pos_iou_thrassigned_gt_inds[pos_inds] = argmax_overlaps[pos_inds] + 1# 4. assign fg: for each gt, proposals with highest IoUfor i in range(num_gts):if gt_max_overlaps[i] >= self.min_pos_iou:if self.gt_max_assign_all:max_iou_inds = overlaps[i, :] == gt_max_overlaps[i]assigned_gt_inds[max_iou_inds] = i + 1else:assigned_gt_inds[gt_argmax_overlaps[i]] = i + 1if gt_labels is not None:assigned_labels = assigned_gt_inds.new_zeros((num_bboxes, ))pos_inds = torch.nonzero(assigned_gt_inds > 0).squeeze()if pos_inds.numel() > 0:assigned_labels[pos_inds] = gt_labels[assigned_gt_inds[pos_inds] - 1]else:assigned_labels = Nonereturn AssignResult(num_gts, assigned_gt_inds, max_overlaps, labels=assigned_labels)

本文来自互联网用户投稿，文章观点仅代表作者本人，不代表本站立场，不承担相关法律责任。如若转载，请注明出处。 如若内容造成侵权/违法违规/事实不符，请点击【内容举报】进行投诉反馈！