# Papers and Articles

Generatic Object Detection
===
- [Object Detection in 20 Years: A Survey (2019)](https://arxiv.org/pdf/1905.05055.pdf)
- [Deep Learning for Generic Object Detection: A Survey (2019)](https://link.springer.com/article/10.1007/s11263-019-01247-4)

Convolutional neural network
===
- [The Understanding of Convolutional Neuron Network Family (2017)](https://pdfs.semanticscholar.org/cd87/9a2eab58e4a1cdcfb9170dbed7df0e948227.pdf?_ga=2.109884098.695306972.1606887373-157373406.1606887373)
- [Visualizing and Understanding Convolutional Networks (2013)](https://arxiv.org/pdf/1311.2901.pdf)

YOLO - You Only Look Once
===
Yolov4
---
- [YOLOv4: Optimal Speed and Accuracy of Object Detection (2020)](https://arxiv.org/pdf/2004.10934.pdf)
- [【AI Meetup】 最強的AI物件偵測技術Yolo-v4作者親自剖析](https://youtu.be/HdQqAF-rMKc)\
王建堯博士從1:03:10開始

Yolov3
---
- [YOLOv3: An Incremental Improvement (2018)](https://arxiv.org/pdf/1804.02767.pdf)

Ensemble Object Detection
===
- [Road Damage Detection using Deep Ensemble
Learning](http://128.84.4.18/pdf/2011.00728)
Using YOLO-v4 as the object detector

### Gridmask Data Augmentation
- https://zhuanlan.zhihu.com/p/103992528

# Mixedup data pre-process 解決周圍背景的影像

- https://zhuanlan.zhihu.com/p/115154110
- https://github.com/ruinmessi/ASFF

# 解決對錯誤目標的高confidence問題
- https://zhuanlan.zhihu.com/p/102870562
- https://medium.com/@jimmyyoung1995/%E8%AB%96%E6%96%87%E7%AD%86%E8%A8%98-learning-con%EF%AC%81dence-for-out-of-distribution-detection-in-neural-networks-a0395d72545f

# vovnet & fixRes
- https://zhuanlan.zhihu.com/p/347537453
- https://github.com/facebookresearch/FixRes
- https://blog.csdn.net/xiaohu2022/article/details/105318534

# Yolov5 vs Yolov4
- https://zhuanlan.zhihu.com/p/161083602
- https://zhuanlan.zhihu.com/p/172121380

<br>
---
<br>

YOLOv4 related issues
===
resolution, distance
---
- [jitter=0 random=0](https://github.com/AlexeyAB/darknet/issues/2268#issuecomment-471491229), detection will be triggered when the camera is placed around 10 meters of the object.
- [random=1](https://github.com/AlexeyAB/darknet/issues/30#issuecomment-280768093), the network resized at different resolutions each 10 iterations (batches)
- [random=0](https://github.com/AlexeyAB/darknet/issues/2593#issuecomment-472515816), If all your Training/Validation/Test images have the same size, then you can train with random=0
- [threshold=.6 above random](https://github.com/AlexeyAB/darknet/issues/30#issuecomment-283706325), I think if you will increase this thresh = .6 then you will get: less false-positives, less true-positives, results with more accurate bounded boxes
- [two ways of resize the resolution](https://github.com/AlexeyAB/darknet/issues/30#issuecomment-283932623), In two cases will be the same accuracy

Negative data
---
[issues 30](https://github.com/AlexeyAB/darknet/issues/30#issuecomment-283782817), you should add images without objects and bounded boxes

Learning Rate
---
[steps=100,1200,2000](https://github.com/AlexeyAB/darknet/issues/30#issuecomment-283923760), VOC has 20 classes, and steps should be divided by 20 for 1 class.

Anchors
---
[Anchor ratio](https://github.com/AlexeyAB/darknet/issues/30#issuecomment-283915206), If we strictly optimize only for apples and only for ratio 4:3

Higher Accuracy
---
[issue 2133](https://github.com/AlexeyAB/darknet/issues/2133#issuecomment-451427102), In theory, to get higher accuracy

CLAHE
---
[Detection in shadows or dark side - issue 4953](https://github.com/AlexeyAB/darknet/issues/4953)

<br>
---
<br>

Topics
===
Data Imbalance
---
- [Imbalance Problems in Object Detection: A Review (2019)](https://arxiv.org/pdf/1909.00169.pdf)

Feature Map
---
- Zeiler and Fergus, 2014 [Visualizing and Understanding
Convolutional Networks](https://cs.nyu.edu/~fergus/papers/zeilerECCV2014.pdf)
- [Adaptive Deconvolutional Networks for Mid and High Level Feature Learning](https://cs.nyu.edu/~fergus/drafts/deconv_iccv_names.pdf)\
提供deconvolution的演算法
- [Feature Visualization](https://distill.pub/2017/feature-visualization/)
- [Feature Visualization with YOLOv3](https://medium.com/@jl22/feature-visualization-with-yolov3-fa133f15cae4)\
Github repo:https://github.com/jennalau/feature-vis-yolov3

NIR for Plastic
---
- [An airborne remote sensing case study of synthetic hydrocarbon detectionusing short wave infrared absorption features identified from marine-harvested macro- and microplastics](https://reader.elsevier.com/reader/sd/pii/S0034425717305722?token=246612C3874309DF5498E3A3A1AD5A1BC2A4CE834E6606039F220F670B784B7439AF8D11BAB052D6009C3F26B3ED9585)
- [Plastic solid waste identification system based on near infraredspectroscopy in combination with support vector machine](https://reader.elsevier.com/reader/sd/pii/S2542504818300113?token=9B43F93C5FE5CB750013E2EC53E1A308B8E2D6A1CA9643E6453A59F05ABEEFFDB616080FC29A4DA0262067AFC7D7C6DF)

CNN on Hyperspectral Data
---
- [Deep Learning for Classification of Hyperspectral Data: A Comparative Review](https://arxiv.org/pdf/1904.10674.pdf)

Database
===
- [Benchmarking of Relational and NoSQL Databases to Determine Constraints for Querying Robot Execution Logs](https://courses.cs.washington.edu/courses/cse544/15wi/projects/Fiannaca_Huang.pdf) \
Evaluates three potential database system (MongoDB, PosgreSQL, SQLite3) against three main criteria:
1. What is the maximum throughput of each database under loads consisting of varying numbers of topics, sizes of messages, and frequency of messages?
PostgreSQL在以不同頻率輸入資料的表現最差，可能原因是PostgreSQL是以extension的方式吃JSON格式資料。
2. Given a real world bag file dataset (specifically, the MIT Stata dataset), how well does the database perform (e.g. percent of messages persisted) when run in a single machine configuration?
在實際資料輸入上SQLite3表現最差，與前一項結果相反，可解釋為是homogeneous和non-homogeneous資料的差異。
3. How rich is the query interface supported by the database, and how fast can typical robotics queries be executed? (e.g. “Select all non-overlapping 1 minute intervals where the robot was in state X”, or “Select all messages beginning 5 minutes before each system error was detected”)
PostgreSQL做indexing並不影響query的速度，因為PostgreSQL並無JSON expression-based indexes。

---

- [A Comparative Study of Databases for Storing Sensor Data](http://www.diva-portal.se/smash/get/diva2:1325707/FULLTEXT01.pdf)
比較對象包含SQL與noSQL:InfluxDB, Elasticsearch, Cassandra, MongoDB, TimescaleDB, and OracleDB
- Section2.1 內的 Data Storage Models 中基於CAP Theorem下又分支為ACID與BASE，適合智慧工廠的情境為**BASE: Basically Available, Soft State, Eventually consistent.** 這樣的Data Storage Model可支援更高的insertion load 和 scalability to multiple machine.
- Section 2.8 Related Work 有對DB相關文獻做summary，比較重要的概念如：
- > It is designed around the
assumptions of single-writer, append-only, and that the most recent data can be retrieved again from the IoT device in the event of a crash
- > delete and update are very rare and thus does not required high performance, ..., when it comes to time series data, no single points of data is too important. In effect, the main focus in on larger data aggregates and not on individual data points.
- time series data 也是智慧工廠會需要的資料情境
- Section 7.1.3 Insights
- 沒有一個是最佳的db for sensor data, Elasticsearch 是 flexibility下適合的選擇
- > If the queries that will be asked of the data is known from the start, if the primary load is of writes or reads, and if the data format might change in the future or stays consistent.
- 智慧工廠的primary load應為**write**, 在未來**data format 可能會變動**

---

- [A Workbench for Quantitative Comparison of Databases in Multi-Robot Applications](http://ropod.org/downloads/IROS18.pdf)
- 比較 neo4j, orientdb, couchdb, mongodb, cassandra, arangodb, influxdb, mysql
- Three typical database architectures:
- Master-slave architecture
- Centralized master-master architecture
- Decentralized master-master architecture
- Conclusion
- MongoDB and ArangoDB performed well
- > the focus of this discussion was on networked multi-robot systems. For these systems the decentralized master-master architecture seems to be the most appropriate
- 但是MongoDB不支援decentralized master-master
- > As a best practice, it is suggested to store blob data files in the file system, along with references in the database.
- [backend版不推薦存blob](https://www.facebook.com/groups/616369245163622/permalink/1147981075335767/)

Ref
---
[MongoDB vs Elasticsearch 比較表](https://leriou.github.io/2019-01-09-mongodb-compareto-elasticsearch/)