---
title: info2022-homework3
tags: info2022
---
# 資訊科技產業專案設計課程 HW3
## 工作職缺
### [Software Engineer, Google Silicon](https://careers.google.com/jobs/results/89995286794904262-software-engineer-silicon/?company=Google&company=YouTube&employment_type=FULL_TIME&gclid=CjwKCAiA9qKbBhAzEiwAS4yeDVWpeoEqdUAbMXMPdEz_IVnr6j43D_AK_ztziZ2JGlj7IV356ZXsURoCKB8QAvD_BwE&gclsrc=aw.ds&hl=en_US&jlo=en_US&location=Taiwan&q=Computer%20Vision,%20Software%20Engineering&sort_by=relevance&src=Online%2FHouse%20Ads%2FBKWS_Cloud_APAC)
:::spoiler

*Minimum qualifications:*
* Bachelor’s degree in Computer Science, Electrical Engineering, related technical field or equivalent practical experience.
* Experience in the field of software engineering (i.e. working with data structures, algorithms, software design).
* Experience in Software development, one or more programming languages such as: C/C++, Java or Python.

*Preferred qualifications:*
* Master's or PhD degree in Computer Science, Electrical Engineering, or related technical field, specializing in computer architecture, operating system, compilers, or hardware-software codesign.
* Experience in compiler design and implementation, image/video/speech algorithms, and software/hardware implementation techniques.
* Experience in programming domain-specific accelerators, including GPU, TPU, or DSP.
* Experience in Android operating system, framework, and application.
* Experience in machine learning software stack and hardware architecture.
* Experience in system power and performance benchmark, profiling, analysis, and optimization.
:::
### [Imaging and Vision Architect, Google Silicon](https://careers.google.com/jobs/results/106748262670574278-imaging-and-vision-architect-devices-and-services-silicon/?%2F=)
:::spoiler
*Minimum qualifications:*
* Master's degree in Electrical Engineering, Computer Science, or equivalent practical experience.
* **5 years** of experience in **digital imaging and computer vision methods and algorithms**.
* **5 years** of experience in **imaging or vision pipeline**.

*Preferred qualifications:*

* Doctorate's degree in Computer Science or Electrical Engineering, or a related field.
* 5 years of relevant industry experience.
* Experience developing hardware IP.
* Experience in **array cameras, computational photography techniques**,**depth sensing cameras**, and others.
* Deep understanding of **machine learning or image/video processing algorithms** for photography applications.
* Understanding of **image quality metrics**.
:::
### [Camera Imaging Software Engineer, Tesla Autonomy](https://www.tesla.com/careers/search/job/camera-imaging-software-engineer-autonomy-89983)
:::spoiler
*Requirements*
* Practical experience programming in C/C++ software, including modern **C/C++** (C++14/17/20), **multithreading**, and **Python**. **Cuda/OpenCL** experience a plus.
* Knowledge of digital camera systems and related concepts, including **color filters, exposure control, timing, drivers, and in-sensor HDR techniques**. Experience with imager bring-up.
* Experience or familiarity with **ISP pipelines**, drivers, and associated tunings
* Experience or familiarity with **Computer Vision, Machine Learning** & related software concepts
* Experience with performant **embedded software design**, object-oriented C++
* Proficient developing software on a Linux host, for embedded Linux targets a plus
:::

### [Machine Learning and Computer Vision Algorithm Engineer, MediaTek](https://careers.mediatek.com/eREC/JobSearch/JobDetail/MTK120160906004?langKey=en-US)
:::spoiler
Requirements
* Familiar with deep learning framework: **Tensorflow, Pytorch, Caffe/Caffe2**.
* Familiar with on-device framework for mobile/embedding devices: **TFLite, ANN**.
* Required programming language: **Python, C/C++, Java**.
* **2+ years experience** **in deep learning** algorithm/model development/optimization, model training methodology.
* Have experience in developing **deep learning algorithms for mobile/embedding devices**.
* Familiar with software development and software engineering, **version control system**
:::

## 專業上匹配程度評估
### Computer Vision and Machine Learning Jobs
* camera 和 depth sensor（LiDAR，Stereo）相關應用實作經驗
* 了解 ISP Pipeline，數位影像處理基礎，經典電腦視覺演算法
* 有 Machine Learning 背景知識會是加分點
* Deep Learning方面開發工具要會用Tensorflow，Pytorch
* Embedded System 實作經驗

### 自我評估
* 非純血電資：中字輩電機，成大AI所
* 具備 Computer Vision 和 Machine Learning 相關背景和實作經驗
* 有把深度學習演算法移植到Embedded System的產學合作經驗
* 自己的研究都有用到camera 和 depth sensor
* 有關ISP的知識要補强
* 缺乏 Multi-Thread Programming 背景
* 缺乏 Android,TPU,CUDA C 等應用開發能力

## 模擬面試
🧔：interviewer 👶：interviewee

🧔：您好，今天會由我來主持這場面試。可以請你先簡單介紹一下自己嗎？
👶：您好我是XXX，目前就讀國立成功大學人工智慧科技碩士學位學程，預計2023年8月畢業。研究領域為**電腦視覺與深度學習**，主要應用在**視覺引導機械手臂控制**上。在學期間曾經把即時手臂控制系統從PC端移植到嵌入式系統上做優化，嵌入式系統的規格都是針對公司產線需求訂立的。過程中除了學到技術外，也學到如何把實驗結果整理好，然後與公司的工程師開會討論系統規格。除了研究以外，我平常也喜歡留意AI技術在其他領域的發展，跟實驗室同學也會討論彼此的研究和看法。

🧔：剛才有提到嵌入式系統上的優化，你可以具體講一下你的做法嗎？
👶：這個手臂控制系統其實已經經過很多前輩開發，程式碼也有幾千行，如果只trace code 的話不太容易發現問題。所以我有利用Vtune 這個profiler來幫助找出效能瓶頸，結果發現程式碼有一部分把灰階影像當成彩色影像來儲存，導致後面利用Gaussian Filter 去除高頻雜訊時多耗費了30 ms。

🧔：請問你是如何判斷可以使用灰階影像來處理問題呢？
👶：彩色影像跟灰階影像最大的分別是，彩色影像保留了物體顔色的資訊，但因為用了3通道（channel），處理彩色影像的運算複雜度會比較高。所以如果我們不需要利用顔色資訊的話就可以用灰階影像作處理，例如車道的綫條偵測就可以用灰階影像作處理。

🧔：那你可以分析一下Gaussian Filter的使用方法嗎？
👶：一般情況Gaussian Filter能有效的去除雜訊，比較大的限制是它不能去除 Salt-and-Pepper Noise。在影像上這種雜訊通常會以極值（0 或 255） 出現，即一個小黑點（0） 或小白點（255）。通常這種雜訊會用 median filter 來處理。

🧔：為何影像會出現Salt-and-Pepper Noise？
👶：通常是 sensor 在 sensing 的過程中，突然被一個
脈衝（pulse）干擾到，很可能就會產生Salt-and-Pepper Noise。
sensor 在 sensing 的過程中，突然被一個脈波（pulse）干擾到，很可能就會產生所謂的 Impulse noise。

🧔：那請問你的手臂控制系統有用到深度學習嗎？
👶：有的，深度學習主要用在辨識物體得到物體的2D影像坐標，然後透過手眼校正的參數把2D坐標轉成3D坐標，最後讓機械手去夾取物體。

🧔：有用到一些深度學習模型優化技術嗎？
👶：我有把深度學習模型從FP32量化成INT8來表示，從而減少模型用到的儲存空間和推論時内存的消耗量。


🧔：你還瞭解其他模型優化技術嗎？你有聽過AutoML嗎？
👶：有，我之前有研究過Google發表的論文，它提出的MobileNetV3和 EfficientNetV1/V2 都有用到 NAS（Neural Architecture Search）。NAS是一種AutoML的技術,主要是讓機器自己設計類神經網路的架構，不過我們要先定義可以選擇的網路基本元素 (如convolution、batch normalization)、與可以調整的內容 (如kernel size、filter number)。

🧔：好,今天的面試就到這謝謝！