前言

本文简述基于chainer实现centerNet目标检测算法

一、centerNet代码实现

from collections import defaultdict
import chainer
import numpy as np
from chainer import Chain, Variable, reporter
from typing import Callable, Dict
import chainer.functions as F
from core.loss import center_detection_loss
from utils.bbox import resize_bbox
from utils.image import resizedef find_peak(map, x, y):dx = np.array([-1,  0,  1, 0, -1,  1, -1, 0, 1])dy = np.array([-1, -1, -1, 0,  0,  0,  1, 1, 1])while True:nx = np.minimum(np.maximum(x + dx, 0), map.shape[1] - 1)ny = np.minimum(np.maximum(y + dy, 0), map.shape[0] - 1)max_idx = np.argmax(map[ny, nx])if x == nx[max_idx] and y == ny[max_idx]:breakx = nx[max_idx]y = ny[max_idx]return x, yclass CenterDetector(Chain):def __init__(self, base_network_factory: Callable[[Dict[str, int]], Chain], insize, num_classes, downratio=4, dtype=np.float32):super().__init__()self.num_classes = num_classesself.insize = insizeself.downratio = downratioself.dtype = dtypewith self.init_scope():self.base_network = base_network_factory({'hm': num_classes,'wh': 2,'offset': 2,})def forward(self, x):y = self.base_network(x)return ydef predict(self, imgs, k=100, detail=False, output_index=-1):x = []sizes = []for img in imgs:_, H, W = img.shapeimg = self._prepare(img)x.append(self.xp.array(img))sizes.append((H, W))with chainer.using_config('train', False), chainer.function.no_backprop_mode():x = Variable(self.xp.stack(x))output = self.forward(x)[output_index]bboxes = []labels = []scores = []output['hm'] = F.sigmoid(output['hm'])output['hm'].to_cpu()for i in range(len(imgs)):bbox, label, score = self._decode_output(output, i, k)bbox = resize_bbox(bbox, (self.insize, self.insize), sizes[i])bboxes.append(bbox)labels.append(label)scores.append(score)if detail:return bboxes, labels, scores, outputelse:return bboxes, labels, scoresdef _decode_output(self, output, index, k):bboxes = []labels = []scores = []for j in range(self.num_classes):hm = output['hm'].array[index, j]pos_indices = np.argsort(hm.flatten())[::-1][:k]already_visit_peak = defaultdict(lambda: False)for pos_index in pos_indices:x = pos_index % hm.shape[1]y = pos_index // hm.shape[1]peak_x, peak_y = find_peak(hm, x, y)if not already_visit_peak[peak_y, peak_x]:already_visit_peak[peak_y, peak_x] = Trueadjusted_x, adjusted_y, w, h = self._decode_bbox(output, peak_x, peak_y, index)bboxes.append([adjusted_y - h / 2, adjusted_x - w / 2, adjusted_y + h / 2, adjusted_x + w / 2])labels.append(j)scores.append(hm[y, x])scores = np.array(scores)sorted_idx = scores.argsort()[::-1]return np.array(bboxes)[sorted_idx], np.array(labels)[sorted_idx], scores[sorted_idx]def _decode_bbox(self, output, x, y, index):wh = output['wh']wh.to_cpu()offset = output['offset']offset.to_cpu()wh = wh.arrayoffset = offset.arrayreturn (x * self.downratio + offset[index, 0, y, x], y * self.downratio + offset[index, 1, y, x],wh[index, 0, y, x] * self.downratio, wh[index, 1, y, x] * self.downratio)def _prepare(self, img):img = img.astype(self.dtype)img = resize(img, (self.insize, self.insize))return imgclass CenterDetectorTrain(Chain):def __init__(self,center_detector,hm_weight,wh_weight,offest_weight,focial_loss_alpha=2,focial_loss_beta=4,comm=None):super().__init__()self.hm_weight = hm_weightself.wh_weight = wh_weightself.offset_weight = offest_weightself.comm = commwith self.init_scope():self.center_detector = center_detectordef forward(self, **indata):imgs = indata['image']y = self.center_detector(imgs)loss, hm_loss, wh_loss, offset_loss, detail_losses = center_detection_loss(y, indata,self.hm_weight, self.wh_weight, self.offset_weight, comm=self.comm)hm = y[-1]["hm"]hm_mae = F.mean_absolute_error(hm, indata["hm"])reporter.report({'loss': loss,'hm_loss': hm_loss,'hm_pos_loss': detail_losses['hm_pos_loss'],'hm_neg_loss': detail_losses['hm_neg_loss'],'hm_mae': hm_mae,'wh_loss': wh_loss,'offset_loss': offset_loss}, self)return loss

总结

源码：chainer-centerNet

chainer-目标检测-centerNet相关推荐

[目标检测]CenterNet
文章目录背景性能模型结构与输出解析目标框(后处理) 损失函数 heatmap reg wh loss 模型四元素: 背景模型结构与输出后处理损失函数背景目标检测就是识别出图中物体的位 ...
CV Code | 本周新出计算机视觉开源代码汇总（含图像修复、目标检测、医学图像分割、度量学习等）...
点击我爱计算机视觉标星,更快获取CVML新技术 CV君汇总了过去一周计算机视觉领域新出的开源代码,涉及到图像质量评价.图像去雾.图像修复.医学图像分割.目标检测.人脸对齐.度量学习等,其中有多篇来自C ...
ICCV2021 | 单目3D目标检测框架GUPNet解析
作者 | Kwong 编辑 | 汽车人原文链接:https://zhuanlan.zhihu.com/p/545796074 点击下方卡片,关注"自动驾驶之心"公众号 ADAS ...
目标检测系列（八）——CenterNet：Objects as points
文章目录摘要 1. 引言 2. 相关工作 3. 准备阶段 4. objects as points 目标即点 4.1 3D检测 5. 实施细节 7. 总结 8. 代码 8.1 训练 8.2 测试摘 ...
重读 CenterNet，一个在Github有5.2K星标的目标检测算法
本文带领大家重温 Objects as Points 一文,其于2019年4月发布于arXiv,谷歌学术显示目前已有403次引用,Github代码仓库已有5.2K星标,无论在工业界和学术界均有巨大影响 ...
【目标检测】CenterNet
论文地址:https://arxiv.org/pdf/1904.07850.pdf 代码地址:CenterNet: Object detection, 3D detection, and pose e ...
目标检测: 一文读懂 CenterNet (CVPR 2019)
论文:Objects as Points 论文链接:https://arxiv.org/pdf/1904.07850.pdf 论文代码:https://github.com/xingyizhou/Ce ...
CenterNet代码解析-ctdet目标检测
CenterNet(Objects as points)开源代码: https://github.com/xingyizhou/CenterNethttps://github.com/xingyizh ...
基于anchor-free的目标检测算法CenterNet研究
2020.04.18 小记近期一直在MOT算法研究,目前SOTA算法核心还是基于CenterTrack网络,而CenterTrack又是基于anchor-free式的CenterNet网络,Cent ...
CenterNet做2D和3D目标检测
论文Objects as Points 源码GitHub地址 CenterNet是全卷积的神经网络,不需要额外的NMS的后处理,属于one-stage的检测方法. 1.2D目标检测通过预测目标的中心 ...

chainer-目标检测-centerNet

文章目录

前言

一、centerNet代码实现

总结

chainer-目标检测-centerNet相关推荐

最新文章

热门文章