A Robust Non-IoU Alternative to Non-Maxima Suppression in Object Detection

Overview

Confluence: A Robust Non-IoU Alternative to Non-Maxima Suppression in Object Detection

1. 介绍

image

用以替代 NMS,在所有 bbox 中挑选出最优的集合。 NMS 仅考虑了 bbox 的得分,然后根据 IOU 来去除重叠的 bbox。而 Confluence 则是利用曼哈顿距离作为 bbox 之间的重合度,并根据置信度加权的曼哈顿距离还作为最优 bbox 的选择依据。

2. 算法原理

2.1 曼哈顿距离

两点的曼哈顿距离就是坐标值插的 L1 范数:

image

推广到两个 bbox 对的哈曼顿距离则为:

image

该算法便是以曼哈顿距离作为两个 bbox 的重合度,曼哈顿距离小于一定值的的 bbox 则被认为是一个 cluster。

2.2 归一化

因为 bbox 有个各样的 size 和 position,所以直接计算曼哈顿距离就没有可比性,没有标准的度量。所以需要对其进行归一化:

image

2.3 置信度加权曼哈顿距离

NMS在去除重合 bbox 是仅考虑其置信度的高低,Condluence 则同时考虑了曼哈顿距离和置信度,构成一个置信度加权曼哈顿距离:

image

3. 算法实现

image

算法:

(1)针对每个类别挑出属于该类别的 bbox 集合 B

(2)遍历 B 中所有的 bbox bi,并计算 bi 和其他 boox的 曼哈顿距离 p,并归一化

2.1 选出 p < 2 的集合,作为一个 cluster,并计算加权曼哈顿距离 wp。 

2.2 在该 cluster 中挑选出最小的 wp 作为 bi 的 wp。 

(3)遍历完毕后,挑出 wp 最小的 bi 作为最优 bbox,添加进最终结果集合中,并将其从 B 去除

(4)把与最优 bbox 的曼哈顿距离小于阈值 MD 的的 bbox 从 B 中去除

(5)不断重复 (2) - (4),每次都选出一个最优 bbox,知道 B 为空

注意:

(1)原文伪代码第 5 行:optimalConfuence 初始化成一个比较大的值就可以,不一定要是 Ip

(2)原文伪代码第 12 行:应该是 Proximity / si

4. 实验结果

image

5. 代码解析

5.1 YOLOv3/4 的后处理

这个接口可以直接处理 YOLOv3/4 的 yolo 层的输出进行后处理

confluence_process(prediction, conf_thres=0.1, wp_thres=0.6)

支持多标签和单标签,并把数据重组后进行 confluence/NMS 处理

# Detections matrix nx6 (xyxy, conf, cls)
if multi_label:
    i, j = (x[:, 5:] > conf_thres).nonzero().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]

5.2 Confluence 算法

confluence(prediction, class_num, wp_thres=0.6)

给所有目标添加上序号

index = np.arange(0, len(prediction), 1).reshape(-1,1)
infos = np.concatenate((prediction, index), 1)

不同类别单独处理,并遍历所有的剩余目标集合 B,直到集合为空,对应上面伪代码的(1)-(2)

for c in range(class_num):       
    pcs = infos[infos[:, 5] == c]             
    while (len(pcs)):                      
        n = len(pcs)       
        xs = pcs[:, [0, 2]]
        ys = pcs[:, [1, 3]]             
        ps = []        
        # 遍历 pcs,计算每一个box 和其余 box 的 p 值,然后聚类成簇,再根据 wp 挑出 best
        confluence_min = 10000
        best = None
        for i, pc in enumerate(pcs):

计算所有目标与其他目标的曼和顿距离 p 和加权曼哈顿距离 wp,p < 2 的目标作为一个 cluster,其中最小的 wp 作为该 cluster 的 wp

index_other = [j for j in range(n) if j!= i]
x_t = xs[i]
x_t = np.tile(x_t, (n-1, 1))
x_other = xs[index_other]
x_all = np.concatenate((x_t, x_other), 1)
.
.
.
# wp
wp = p / pc[4]
wp = wp[p < 2]

if (len(wp) == 0):
    value = 0
else:
    value = wp.min()

选出最小的 wp,确定目标

# select the bbox which has the smallest wp as the best bbox
if (value < confluence_min):
   confluence_min = value
   best = i  

然后把与目标的曼哈顿距离小于阈值的目标和本身都从集合 B 中去除

keep.append(int(pcs[best][6])) 
if (len(ps) > 0):               
    p = ps[best]
    index_ = np.where(p < wp_thres)[0]
    index_ = [i if i < best else i +1 for i in index_]
else:
    index_ = []
    
# delect the bboxes whose Manhattan Distance is below the predefined MD
index_eff = [j for j in range(n) if (j != best and j not in index_)]            
pcs = pcs[index_eff]

最后继续重复遍历集合 B,直到集合为空。

仓库里我放了一张测试照片和原始检测结果,大家可以直接用来调试 confluence 函数。

Credits:

https://arxiv.org/pdf/2012.00257.pdf

A collection of Jupyter notebooks to play with NVIDIA's StyleGAN3 and OpenAI's CLIP for a text-based guided image generation.

StyleGAN3 CLIP-based guidance StyleGAN3 + CLIP StyleGAN3 + inversion + CLIP This repo is a collection of Jupyter notebooks made to easily play with St

Eugenio Herrera 176 Dec 30, 2022
Trying to understand alias-free-gan.

alias-free-gan-explanation Trying to understand alias-free-gan in my own way. [Chinese Version 中文版本] CC-BY-4.0 License. Tzu-Heng Lin motivation of thi

Tzu-Heng Lin 12 Mar 17, 2022
The FIRST GANs-based omics-to-omics translation framework

OmiTrans Please also have a look at our multi-omics multi-task DL freamwork 👀 : OmiEmbed The FIRST GANs-based omics-to-omics translation framework Xi

Xiaoyu Zhang 6 Dec 14, 2022
An alarm clock coded in Python 3 with Tkinter

Tkinter-Alarm-Clock An alarm clock coded in Python 3 with Tkinter. Run python3 Tkinter Alarm Clock.py in a terminal if you have Python 3. NOTE: This p

CodeMaster7000 1 Dec 25, 2021
A higher performance pytorch implementation of DeepLab V3 Plus(DeepLab v3+)

A Higher Performance Pytorch Implementation of DeepLab V3 Plus Introduction This repo is an (re-)implementation of Encoder-Decoder with Atrous Separab

linhua 326 Nov 22, 2022
S2s2net - Sentinel-2 Super-Resolution Segmentation Network

S2S2Net Sentinel-2 Super-Resolution Segmentation Network Getting started Install

Wei Ji 10 Nov 10, 2022
Readings for "A Unified View of Relational Deep Learning for Polypharmacy Side Effect, Combination Therapy, and Drug-Drug Interaction Prediction."

Polypharmacy - DDI - Synergy Survey The Survey Paper This repository accompanies our survey paper A Unified View of Relational Deep Learning for Polyp

AstraZeneca 79 Jan 05, 2023
Implementation of Transformer in Transformer, pixel level attention paired with patch level attention for image classification, in Pytorch

Transformer in Transformer Implementation of Transformer in Transformer, pixel level attention paired with patch level attention for image c

Phil Wang 272 Dec 23, 2022
Code for "Typilus: Neural Type Hints" PLDI 2020

Typilus A deep learning algorithm for predicting types in Python. Please find a preprint here. This repository contains its implementation (src/) and

47 Nov 08, 2022
Categorizing comments on YouTube into different categories.

Youtube Comments Categorization This repo is for categorizing comments on a youtube video into different categories. negative (grievances, complaints,

Rhitik 5 Nov 26, 2022
ImageNet-CoG is a benchmark for concept generalization. It provides a full evaluation framework for pre-trained visual representations which measure how well they generalize to unseen concepts.

The ImageNet-CoG Benchmark Project Website Paper (arXiv) Code repository for the ImageNet-CoG Benchmark introduced in the paper "Concept Generalizatio

NAVER 23 Oct 09, 2022
[WWW 2022] Zero-Shot Stance Detection via Contrastive Learning

PT-HCL for Zero-Shot Stance Detection The code of this repository is constantly being updated... Please look forward to it! Introduction This reposito

Akuchi 12 Dec 21, 2022
Code for the paper "Regularizing Variational Autoencoder with Diversity and Uncertainty Awareness"

DU-VAE This is the pytorch implementation of the paper "Regularizing Variational Autoencoder with Diversity and Uncertainty Awareness" Acknowledgement

Dazhong Shen 4 Oct 19, 2022
An implementation of Geoffrey Hinton's paper "How to represent part-whole hierarchies in a neural network" in Pytorch.

GLOM An implementation of Geoffrey Hinton's paper "How to represent part-whole hierarchies in a neural network" for MNIST Dataset. To understand this

50 Oct 19, 2022
Backdoor Attack through Frequency Domain

Backdoor Attack through Frequency Domain DEPENDENCIES python==3.8.3 numpy==1.19.4 tensorflow==2.4.0 opencv==4.5.1 idx2numpy==1.2.3 pytorch==1.7.0 Data

5 Jun 18, 2022
An Api for Emotion recognition.

PLAYEMO Playemo was built from the ground-up with Flask, a python tool that makes it easy for developers to build APIs. Use Cases Is Python your langu

greek geek 2 Jul 16, 2022
(CVPR2021) DANNet: A One-Stage Domain Adaptation Network for Unsupervised Nighttime Semantic Segmentation

DANNet: A One-Stage Domain Adaptation Network for Unsupervised Nighttime Semantic Segmentation CVPR2021(oral) [arxiv] Requirements python3.7 pytorch==

W-zx-Y 85 Dec 07, 2022
This repository contains the implementation of the paper: "Towards Frequency-Based Explanation for Robust CNN"

RobustFreqCNN About This repository contains the implementation of the paper "Towards Frequency-Based Explanation for Robust CNN" arxiv. It primarly d

Sarosij Bose 2 Jan 23, 2022
Fibonacci Method Gradient Descent

An implementation of the Fibonacci method for gradient descent, featuring a TKinter GUI for inputting the function / parameters to be examined and a matplotlib plot of the function and results.

Emma 1 Jan 28, 2022
Torch implementation of SegNet and deconvolutional network

Torch implementation of SegNet and deconvolutional network

Fedor Chervinskii 5 Jul 17, 2020