Today Finish:

  1. Review Paper No-Reference Quality Assessment for Screen Content Images Based on Hybrid Region Features Fusion

    (1) SEPes:

    SEPes: text, computer graphics and high-frequency part of pictorial region in SCI

    Non-SEPes: smooth and background areas with noise (the distribution of non-SEP is uniform, i.e. the intensity of image changes slowly, and the distortions are smoothed → relatively small effect)

    A. Sharp edge patches (SEPes) <- variance of local standard deviation

    B. Feature Extraction:

    entropy and contrast features from a gray-level co-occurrence matrix → local phase coherence

    local phase coherence → capture the loss in sharpness

    C. Average pooling → fuse features obtained from all of the SEPes to represent the local features

    D. BRISQUE method → combine local features with global features → hybrid region (HR)-based features

    (2) Some Terminologies:

    GLCM: Gray-level Co-occurrence Matrix

    LPC: local phase coherence

    LSD Distribution: local standard deviation distribution (The LSD indicates the image structural complexity in a local area, which also highlights the details of the edge and smoothes out the effects of distortion.)

    (After being contaminated by severe Gaussian noise, the characteristic of the LSD distribution of the SEP and non-SEP can still be roughly maintained)

    VOLSD: variance of local standard deviation

    implement the classification of SCIs regions

    (3) Feature Extraction:

    A. Motivation:

    a. The analysis of SCIs microstructure is very essential

    b. The texture is concerned as the spatial distribution and spatial dependence among the graylevel in a local area.

    (4) Some Characteristics & Properties of SCIs:

    A. Different regions of SCIs present dissimilar peculiarity

    B. HVS is highly sensitive to image contents containing sharp edges,

    C. Sharp edges are common part of SCIs

    (5) MSCN Histogram:

    BRISQUE first estimates parameters of the presumed generalized Gaussian distribution (GGD) or asymmetric generalized Gaussian distribution (AGGD) by fitting histogram of mean subtracted contrast normalized (MSCN) coefficients or pairwise products of MSCN coefficients from neighboring pixels.

    (6) Some Conclusions:

    A. Influence of Gaussian Window Size

    7 X 7 Gaussian window size

    B. Influence of Patches Number

    More patches, better performance

    set 100 patches

    Different image contents are contained in the same patch create the impressive influences when features extraction are conducted.

  1. Read Paper Blindly Assess Image Quality in the Wild Guided by A Self-Adaptive Hyper Network

(1) Authentically Distorted Images Properties:

A. distortion diversity

B. content variation

C. local distortions

(2) Authentic Distortions:

A. Global uniform distortions (e.g. out of focus, low illumination)

B. Contain other kinds of non-uniform distortions (e.g. object moving, over lighting, ghosting) in local areas

(3) Authentic IQA database:

A. LIVE Challenge (LIVEC)

LIVEC contains 1162 images taken from different photographers with varies camera devices in the real world, hence these images contain complex and composite distortions.

B. KonIQ-10k

KonIQ-10k consists of 10073 images which are selected from the large public multimedia database YFCC100m, the sampled images try to cover a wide and uniform quality distribution in the sense of brightness, colorfulness, contrast and sharpness.


BID is a blur image database containing 586 images with realistic blur distortions such as motion blur and out of focus, etc.

(4) Synthetic IQA Field:

A. hand-crafted feature based IQA - NSS models to capture distortion

(a) Quality aware natural scene parameters include discrete wavelet coefficients, the correlation coefficients across subbands, DCT coefficients, locally normalized luminance coefficients with their pairwise products, image gradient, log-Gabor responses and color statistics.

(b) Distribution models used to capture the statistics from synthetically distorted image include generalized Gaussian distribution (GGD), asymmetric generalized Gaussian distribution (AGGD), Weibull distribution, third order polynomial tting and histogram counting.

B. learning feature based IQA

(a) Codebook based learning approach

(b) CNN based methods

Semantic Features from the Pretrained baseline network: VGG-16, ResNet-50, ResNet-34


<1> ignore how image semantics influence quality perception as mixing semantic learning and quality prediction in one network

<2> local distortions are ignored as deep semantic features are extracted from global scale

(5) Experiments

A. Single database evaluation

B. Generalization ability test (cross database test)

Tomorrow Work:

  1. Review SC Images Codes
  2. Continue to read the above paper


Today Finish:

  1. Problem of current IQA methods for SCIs:

    (1) Divided large SCIs into image patches for data augmentation.

    A. A single image patch can not represent the quality of the entire image, especially in IQA of SCI.

    B. SCI patches of an entire image degraded by the same distortion type and strength may have drastically different quality.

    (2) Assign differential mean opinion score (DMOS) to all image patches

    Inherent error and inaccuracy between DMOS and quality scores of image patches.

    (3) Adopted mean square error (MSE) between the predicted quality and the subjective differential mean opinion score (DMOS), without considering quality ranking between different SCIs. (The main reason is that they do not consider the ranking information in term of quality between SCIs. → small MSE error but poor performance)

    (4) Most databases are laboratory generated, i.e., synthetically distorted images. (authentic distortions)

  2. Patches Size of SCIs:

    Pre-processing: The SCIs are first divided into multiple regions averagely, and an image patch of size 128 × 128 is extracted from each region as a representative sample of each region.

    (1) 128 X 128

    For Natural IQA models, 32 X 32 is the most optimum size.

  3. Ideas of some current methods:

    (a) Assess the quality of textual regions and pictorial regions separately.

    (b) (Better image quality representation) Obtain multiple features, e.g., local, global perceptual features/representations.

    (Multi-region) local features from textual regions and local features from pictorial regions

    global features from the entire image

    (c) Multi-task training: quality score prediction, distortion type (noise) classification

    (d) Different ranking loss to rank image


    1. Local Feature Extraction Module:

      A. The input SCIs are divided into multiple regions, and one image patch of size 128 × 128 is extracted from each region as a representative sample of each region.

      B. In the VGG-based network, combination of two 3 × 3 convolutional layers and one pooling layers owns a larger view to extract local features with less data, which has an excellent feature extraction ability.

      C. The first two convolutional layers adopt a large-scale convolutional kernel of size 5 × 5 to acquire general information of local features from an entire input patch.

    2. Pseudo Global Feature Generation Module:

      A. Fuses local features to generate pseudo global features.

      B. The concatenate layer is used to fuse features of multiple regions,

      C. 1 × 1 convolutional layer is utilized for feature fusion and feature dimension reduction.

    3. Multi-Task Training Module:

      A. Pseudo global features are used to train a multi-task learning model.

      B. Noise classification task and quality score prediction task.

      C. Noise classification: two fully connected (FC) layers and one softmax layer. Since the image subjective quality depends on noise type, noise strength and image content, this branch adopts a classification network to extract features of noise type and strength.

      D. Quality score prediction: three FC layers and one concatenate layer.

      E. Feature vectors of noise classification task and quality score prediction task are concatenated into a new feature vector of quality score prediction task.

    4. Siamese Network Module:

      A. Extract features of different SCIs with shared weights of the proposed model.

      B. Two different SCIs are input into one model simultaneously, and then two predicted scores are obtained as output of the model.

    5. *The advantages of utilizing multi-region features:*

      A. Compared to local features of image patches, the pseudo global features are better representation of the entire image quality, and are reasonable to be labeled with DOMSs;

      B. Compared to using large image patches, utilizing multi-region image patches can obtain more training samples to solve the problem of insufficient data, and take into account the characteristics of the entire image, which has less computation for employing the shared local feature extraction module and a 1 × 1 convolutional layer;

      C. Utilizing multi-region features can reduce the influence of image patch contents with large naturalness statistical differences between SCI patches.

    6. *Loss Function:*

      (The noise loss of the noise classification task adopts the empirical cross entropy loss which shows superior performance in classification models.)

      (The smooth L 1 loss is less sensitive to outliers than the L2 loss, and owns better fitting ability than L1 loss.)

      (This ranking loss can make networks learn the quality difference of two different SCIs resulting in owning ranking ability.)

    7. Quality Score:

      DMOS values are ranged from 0 to 100, where 0 indicates the best quality, and 100 indicates the worst quality.

    8. Evaluation Criteria:

      o_i and s_i are the objective and subject scores

      e_i is the difference between the subjective and object results

      A. Pearson Linear Correlation Coefficient (PLCC)

      measure a method’s prediction accuracy

      B. Spearman Rank Order Correlation Coefficient (SRCC)

      measure a method’s prediction monotonicity

      Non-parametric rank-order based correlation metric that is independent of any monotonic score mapping.

      It is employed to access prediction monotonicity

      C. Kendalls Rank-order Correlation Coefficient (KRCC)

      measure for monotonicity prediction

      D. Root Mean Square Error (RMSE)

      RMSE can be adopted to gauge the prediction consistency.

      E. Statistical Significance

      signify whether the difference in the performance of one IQA method with respect to another, on a set of sample points, is purely due to chance or due to some genuine underlying effect

      F. Five-parameter mapping function

      Nonlinearly regress the quality scores into a common space

      Objective quality scores of SCIs may have different ranges. It is necessary to map the above-mentioned scores into a common range

      x: score estimated by the proposed model

      Q(x): corresponding mapped score

      β1 ,. . . , β5: parameters to be computed with a curve fitting process (fitted by minimizing the sum of squared errors)

Tomorrow Work:

  1. Review another paper again
  2. Continue to review SC IQA Codes


May 14, 2021

Today Finish:

  1. Summarize the following two papers:
    A. No-reference screen content image quality assessment based on multi-region features

    B. No-Reference Quality Assessment for Screen Content Images Based on Hybrid Region Features Fusion

    (1) Screen Content images (SCIs): mixed contents including textual and pictorial regions, e.g., natural scenes pictures, documents and texts images, and computer-generated graphics.

    (2) SCIs Properties: sharp edges, thin lines, little color variations for massive existence of texts and computer-generated graphics.

    (3) Natural Images (NIs) Properties: Continuous-tone content with smooth edges, thick lines, textures, rich color changes

    (4) SCIs and NIs Differences:

    A. SCIs have more sharp edges → Features change significantly

    B. SCIs appear dissimilar treads of variations (dissimilar peculiarity) when they are contaminated by various kinds of distortions with different intensities, and the content of different regions varies greatly.

    Notice: Mean Subtracted Contrast Normalized (MSCN) coefficients of pictorial and textual regions (the MSCN Histogram)

    “The MSCN coefficient histogram of the pictorial region exhibits a Gaussian-like appearance. By contrast, the textual region yields a quite different MSCN distribution.”

    (5) (Model) Input Images:

    A. resized image

    B. image patch, e.g., size 24 X 24, 32 X 32 (overlap or non-overlap)

    (6) Benchmark Dataset: screen content image quality assessment database (SIQAD)

    current SOTA (SRCC): ~ 0.852

    Our Goal (SRCC): ~ 0.91 - 0.95

    The SIQAD is built for evaluating the perceptual quality of the SCIs, which consists of 20 pristine SCIs and 980 corresponding images distorted by 7 types of distortions on 7 distortion levels, involving Gaussian noise (GN), Gaussian blur (GB), motion blur (MB), contrast change (CC), JPEG compression (JPEG), JPEG2000 compression (JP2K) and layer segmentation-based coding (LSC).

    We run 10 times of this random train-test splitting operation and the median SRCC and PLCC values are reported.

    (7) Dataset Partition: 60% Training, 20% Validation, 20% Testing

  2. Review SSIM Paper: Image Quality Assessment: From Error Visibility to Structural Similarity

    (1) objective image quality metric Applications:

    A. It can be used to dynamically monitor and adjust image quality.

    B. It can be used to optimize algorithms and parameter settings of image processing systems (i.e., use this IQA Metric as the object / loss function for the following low-level vision tasks: image enhancement, image denoising, blind image deblurring, image super-resolution, lossy image compression, and image generation, and etc.).

    C. It can be used to benchmark image processing systems and algorithms.

    (2) Why does design an better IQA metric an **optimization** task / problem?:

    as shown above objective image quality metric Applications

    1. **Optimize** a better IQA metric can evaluate and assess the quality 
    of distorted images.

    2. It can be used to ***optimize** algorithms and parameter settings
    of image processing systems*
    ***(i.e., use this IQA Metric as the object / loss function
    for the following low-level vision tasks:
    image enhancement, image denoising, blind image deblurring,
    image super-resolution, lossy image compression,
    and image generation, and etc.).***

    , and

    Our primary goal for designing a better IQA metric is

    1. **Fidelity**: keep the contents of the distorted images 
    (semantic information) unchanged
    2. **Quality**: try to improve the quality for low-level vision tasks

    (3) Objective IQA Types:

    A. Full-reference (FR): have a complete reference image

    B. No-reference (NR) or “blind” IQA: the reference image is not available

    C. Reduced-reference (RR): the reference image is only partially available as a set of extracted features (statistics)

    (4) Widely used IQA algorithms:

    A. Mean Squared Error

    Pros: ideal target for optimization

    a. based on a valid distance metric ($L_2$)

    b. satisfy positive definiteness

    c. symmetry

    d. triangular inequality properties

    e. convex

    f. differentiable

    g. memoryless

    h. additive for independent sources of distortions

    i. energy preserving under orthogonal or unitary transformations


    a. poor correlation with perceptual image quality

    b. based on point-wise signal differences, which are independent of the underlying signal structure.

    Some Thought between MSE and SSIM:

    The simplest implementation of this concept is the MSE, which objectively quantifies the strength of the error signal. But two distorted images with the same MSE may have very different types of errors, some of which are much more visible than others.

    *Reference patch 和 Distorted patch 的相减squared后sum如果相等的话,就认为MSE就相等。但是”But two distorted images with the same MSE may have very different types of errors, some of which are much more visible than others.“ 但SSIM 告诉我们:A和A’、A’’、A’’’等等的它的distorted images的distance不仅仅和它们之间的距离有关,而且和他们本身的特性(比如说图像的object的structure)有关。*

    (最本质理解图像质量) A和B是不同content,虽然A到A‘之间的距离相等,B到B’之间的距离相等,但是它们的perceptual content不同,所以叫做adaptive to local content

    FYI, what is structure distortions and texture resampling?

    structure distortions of an image: artifacts due to noise, blur, or compression

    texture resampling: exchanging a texture region with a new sample

    (5) Generic IQA Framework:

    (6) Hypothesis of the SSIM: Human Visual System (HVS) is highly adapted for extracting structural information.

    Natural Images’ pixels exhibit strong dependencies (spatially proximate) → dependencies carry important information about the structure of the objects in the visual scene.

    (7) SSIM Philosophy V.S. Error Sensitivity Philosophy:

    A. Quantify Image degradation: perceived changes in structural information variation V.S. perceived errors

    B. Paradigm: top-down approach V.S. bottom-up approach

    C. Natural image complexity and de-correlation problem: structural changes V.S. accumulating the errors

    (8) How to calculate the SSIM Index?:

    A. Independent three components: luminance, contrast, and structure

    B. luminance: mean intensity

    (The above formula is qualitatively consistent with Weber’s law, which has been widely used to model light adaptation (also called luminance masking) in the HVS)

    (*the HVS is sensitive to the relative luminance change, and not the absolute luminance change.*)

    C. Contrast: standard deviation

    (With the same amount of contrast change , this measure is less sensitive to the case of high base contrast than low base contrast. This is consistent with the contrast-masking feature of the HVS.)

    D. Structure: normalized signals (unit standard deviation)


    1. Unit vectors each lying in the hyperplane → indicating the structures of the two images
    2. The correlation (inner product) between these is a simple and effective measure to quantify the structural similarity.

    The structure comparison formula:

    E: Overall Similarity Measure:

    (9) Relationships between the SSIM and other metrics:

    Equal-distortion contours drawn around three different example reference vectors, each of which represents the local content of one reference image.

    Each contour represents a set of images with equal distortions relative to the enclosed reference image.

    "**adaptive to local content"**

    illustrated geometrically in a vector space of image components (e.g., pixel intensities, extracted features, or transformed linear coefficients)

    shape: depends on the metric formula, e.g., MSE→circle

    size: Even though different local contents keep equal distortion, the contour may show different sizes due to signal magnitude

    (a) Minkowski metric (Assumed MSE, an exponent of 2): circle, each contour has the same size and shape → perceptual distance corresponds to Euclidean distance

    (b) Minkowski metric (Assumed different image components are weighted differently using CSF): ellipses, each contour has the same size

    (c) Adaptive distortion metric: rescaling the equal-distortion contours according to the signal magnitude

    (d) Contrast masking (magnitude weighting) combination followed by component weighting

    (e) SSIM: separately computes a comparison of two independent quantities: the vector lengths, and their angles. And the contours will be aligned with the axes of a polar coordinate system

    (f) SSIM: computed with different exponents compared with (e)

    (this may be viewed as an adaptive distortion metric, but unlike previous models, both the size and the shape of the contours are adapted to the underlying signal.)

    (10) Why use local patch? ****

    e.g., 8 X 8 square window and moves pixel-by-pixel over the entire image. —> undesirable “blocking” artifacts

    SSIM: 11 X 11 circular-symmetric Gaussian weighting function —> locally isotropic property

    A. luminance and contrast can vary across a scene

    B. image statistical features are usually highly spatially non-stationary

    C. image distortions may also be space-variant

    D. only a local area in the image can be perceived with high resolution by the human observer at one time instance

    E. localized quality measurement can provide a spatially varying quality map of the image, which delivers more information about the quality degradation of the image and may be useful in some applications.

    As a result, for the SSIM index, they used spatial patches extracted from each image.

    One Problem: the input patch and distorted patch must be perfectly aligned with each other

    Another wait-to-discuss problem: Is it really great to compare the quality of patches, or can we compare the full image directly rather than the patches?

  3. Review SC IQA Codes

Tomorrow Work:

  1. Continue to review SC IQA Codes
  2. connect to the server and download the dataset
  3. Run the codes and get results



读paper “A Statistical Evaluation of Recent Full Reference Image Quality Assessment Algorithms”部分内容的总结:

  1. IQA Goal: automatically assess the quality of image or videos in agreement with human quality judgement

  2. Source Image Content: resized images using bicubic interpolation

  3. Image Distortion Causes: impairments such as transmission errors and depend on the image compression scheme used.

  4. Image Distortion Types:
    1> JPEG2000 Compression: generated by compressing the reference images (full color) using JPEG2000 at bit rates ranging from 0.28 bits per pixel (bpp) to 3.15 bpp.

    2> JPEG Compression: at bit rates ranging from 0.15 bpp to 3.34 bpp.

    3> White Noise: White Gaussian noise of standard deviation 𝜎 (Value between 0.012 and 2.0) was added to R, G, B components. And the distorted components were clipped between 0 and 1, and rescaled to the range 0 to 255.

    4> Gaussian Blur: R, G, B components were filtered using a circle-symmetric 2D Gaussian kernel of standard deviation 𝜎 (0.42 to 15 pixels).

    5> Simulated fast fading Rayleigh (Wireless) channel: Images were distorted by bit errors during transmission of compressed JPEG2000 bitstream over a simulated wireless channel.

    6> Smoothness: Edge smoothness or texture blur (the received image is smoother than the original). Loss of high frequency components when compared with the original image.

    7> Blocking: appear in all block-based compression techniques. due to coarse quantization of frequency components

    observed as surface discontinuity (edge) at block boundaries, and the edges are perceived as abnormal high frequency components in the spectrum

    8> Ringing: periodic pseudo-edges around original edges, due to improper truncation of high frequency components

    9> Masking: reduction in the visibility of one image component (target), due to the presence of another (the masker)

     A. luminance masking (light adaptation)
     B. texture masking (occurs when maskers are complex textures or masker and target have similar frequencies and orientations)

    10> Lost block: an alteration of a pixel value, so that it doesn’t match with its neighbors pixel value.


Image Quality Assessment - SSIM.pdf


SSIM Figure 圈为什么是圆圈?


Distorted image 和 中心image 距离一样 contour也是一样


重新review下宝亮师兄之前让我看过的2篇Screen Contents IQA 2篇论文


Read 10 papers using traditional methods and 5 papers using deep learning methods

  1. SSIM paper的Figure 4为什么是圈?

  2. 问题1:SSIM比MSE好在哪?MSE的问题是什么?(很基础的问题)

    MSE 相减后 squared后sum

    Reference patch 和 Distorted patch 的(相减后 squared后sum 相等) 就认为MSE就相等。

    但SSIM 告诉我们:A和A’ distance不仅仅和它们之间的距离有关,而且和他们本身的特性有关

    (最本质理解图像质量)A和B是不同content,虽然A到A‘之间的距离相等,B到B’之间的距离相等,但是它们的perceptual content不同,

    所以叫做adaptive to local content

    问题2:Feature Domain的SSIM —> DISIS

  3. 问题3:SSIM代码:local patch的mean和variance :7X7, 9X9, 11X11

    为什么算local的structural similarity?

  4. No reference说Natural Scene statistics (NSS) 但是 Full Reference不说NSS

  5. 做完了SSIM后,VGG domain (Feature Domain)能不能算SSIM 算的话 代表什么物理意义?

  6. VGG做完, 在做Transformer? 基础的东西,得搞清楚。Deep Learning Features别人没有做的实验 咋们要做 try下 再做fancy的东西

  7. 每天工作12小时,不会说这一天没有汇报的东西 每周开会 “老师 我看到它,它为什么这么做,我是怎么理解想这个问题的”

  8. “为什么这个东西适合graph 为什么适合Transformer?“

  9. 在CityU做研究,而不是工程性的东西 提出问题(他为什么这么做), try our best去解答

  10. 把宝亮师兄想要做的想明白,然后按照师兄的想法去做,思路好,结果可能好;思路不好,结果可能不好。(什么正确的做实验的方式)

  11. 看到efforts,保证每天12个小时工作时间。

  12. 做别人替代不了的东西!打基础 踩坑 进去到领域。

Graduate School in Hong Kong Part I


Decision in 2020 (Senior Year at College)

During my senior year, I decided to attend a graduate school in Australia, majoring in Computer Science and Technology. So, I applied for The Group of Eight (Go8), including The University of Sydney (USYD), University of New South Wales (UNSW), Monash University, University of Adelaide, and University of Technology Sydney (UTS). USYD, however, is the best among all these universities. The reasons why I chose to attend USYD are as follows. First of all, it ranks top 40 in the QS World University Rankings in 2021, listed in the Top 100 World Universities inside the Chinese companies, universities, and government. It is of great benefit to hunt for a job after graduation, and I won’t be restricted by the university that not being ranked high enough. Besides, USYD is quite famous in mainland China, and almost all the Chinese know USYD well due to the city of Sydney. So, If I attended USYD, I could obtain a better (fancy) education background in the resume, which is beneficial for job promotion and salary raising in the future. Furthermore, I’d have an opportunity to experience the culture & society of a foreign country, and learn the English language as well.

USYD Campus, taken by one of my friends

However, in 2020, no one can expect the coming of the COVID-19 pandemic. The COVID virus quickly swept the world and infected millions of people. The borders of western countries started to close and were blocked. Countries announced strict restrictions, and no one except citizens can enter the country. As a result, almost all my friends, who decided to study abroad, started their Bachelor’s, Master’s, or Ph.D.’s Program online and work from home (WFH). BTW, one of my friends told me that lots of employees at giant companies were starting to work from anywhere (WFA), i.e., forever work from home. My thought was to wait for the ending of the pandemic. Meanwhile, I went to take internships in China (Philips and Tencent). After almost half a year, unexpected UK and South Africa variants dominated the infection and killing. Consequently, the border of foreign countries kept closed with no sign of reopening, and I have to make another decision for my future.

Global COVID-19 Pandemic Map, date up to April 9th, 2021

Decision in 2021 (One Year Gap after graduation)

My primary choice was to stay in mainland China. I applied for the M.Phil. in Computer Science at The Chinese University of Hong Kong (CUHK Shenzhen). Unfortunately, I was rejected because of poor English grades, i.e., my IELTS Test Grade cannot meet the university’s minimum requirement. Another choice, after consideration, was to go to Hong Kong, where the pandemic was not severe, and the city was protected from massive infections well. Thus, I applied for the Master of Science CS Program (One-year Program) in Hong Kong: The University of Hong Kong (Rejected), The Hong Kong University of Science and Technology (Rejected), The Hong Kong Polytechnic University (Rejected), Hong Kong Baptist University (Firm Offer). I also applied for the Master of Philosophy CS Program (Two-year Program): City University of Hong Kong (Firm Offer), The Hong Kong Polytechnic University (Not submit due to poor English grade). They said that the application procedures were highly competitive since most Chinese outstanding students cannot go to the USA, Europe, Canada, etc., due to the pandemic. They all applied to universities in HK and Singapore. After discussing with my parents, friends, and professors, they all supported me to attend City University of Hong Kong (CityU HK) for my Master’s Study.

Kowloon Tong, Hong Kong, China

There are multiple pros why CityU HK is a better place to attend:

  1. CityU HK ranked top 50 globally (QS) -> listed Top 100 World Universities recognized by the Chinese government.
  2. After a two-year Master of Philosophy CS, I can apply for a CS Ph.D. position if I could publish papers.
  3. CityU is located in a metropolis, i.e., Hong Kong. I could experience capitalist society in China without studying/going abroad.
  4. Less Tuition Fees -> CityU: 70,000 RMB V.S. USYD: 47,500 AUD (~ 237,500 RMB) per year
  5. Full English Teaching

Cons are also listed below:

  1. No English living environment, and must study and speak Cantonese.
  2. Small Accommodation Area (~ 5 m2 per person).

Skyscrapers in Hong Kong, China

Study and Stay VISA Application

After receiving the Firm Offer of M.Phil. CS from CityU, I applied for the Student Label to enter Hong Kong through CityU Student Visa/Entry Permit Sponsorship. For around one month (March 3 - 27), the Study VISA Label has been approved, and it was sent to me on April 1. Then, I went to the Exit-Entry Administration Bureau for the Stay Visa (D Type VISA). Three working days later, I got the stay visa successfully. Meanwhile, I’m also searching for a residence, e.g., studio flat, in Kowloon, HK. Since the class 2021 students haven’t graduated yet, very few apartments and rooms were available. As a result, I have to rent a single room with a separate bathroom (5,500 HKD per month) in Yau Ma Tei (油麻地), a little far from the university and would spend around 30 mins taking the subway to arrive CityU. I also reserved a double room (4,800 HKD) in Parc Oasis (又一居), and I can move there in Mid-June. The residence is quite near CityU, which takes no more than 10 mins. It’s super convenient.

Google Map Direction from Parc Oasis to CityU

When people take planes to the Hong Kong International Airport, they have to be quarantined for at least 14 days in a hotel, which you know, quite costly. Since I have rented an apartment and wanted to be quarantined at home to save money, I decided to fly to Shenzhen, China first, then enter HK through Shenzhen Bay Port (深圳湾口岸). In this way, I can take a taxi to my apartment and prepare food, drinking water, daily necessities, etc.

Passengers Traveling To and From Shenzhen Bay Port by Taxi

Finally, I arrived at the Pitt Street in Yau Ma Tei! I’m thrilled. However, I found that the tenement buildings are quite dilapidated and old than mainland China apartments. And the surroundings are noisy and dirty. I can even hear whistles and noises of a car and whispers of passer-by! Jesus! What a terrible place! Some people also say the area is not safe. For God’s sake, my suggestion is that you do not rent an apartment in Yau Ma Tei! And if you wanna feel comfortable, please move to a high-rise building in the neighborhood (community). Anyway, I have to stay here for a moment (maybe two months, I think). Another essential issue is a 14-day quarantine at home for inbound travellers. I can go nowhere but stay at home for 14 days and be responsible for the locals even though I have been vaccinated in March. That’s depressing.

Yau Ma Tei, HK

Yau Ma Tei, HK

What I observed are as follows:

  1. Almost all the people in HK speak Cantonese, but they all understand Mandarin Chinese. We can talk in Mandarin with them, and they might reply to us in Mandarin or Cantonese. And, They seldom communicate with each other in English.

Yau Ma Tei, HK

  1. In mainland China, people use online payment like WeChat Pay and Alipay. In contrast, in HK, the locals prefer to pay in HKD cash.

Paying using Cash (HKD) in HK

  1. In HK, most goods in the supermarkets are foreign imports. The illustrations and specifications of them are all in English.

Some snacks in HK

Indeed, I’m looking forward to going to CityU after the quarantine.

Last Update: Apr 17th, 2021

Find Meaning in the Little Moments that Make up Our Life

About me

My name is 贾舒越 (Shuyue Jia). I like to be called Bruce as well. I’m currently an M.Phil. CS Student at City University of Hong Kong (CityU). Earlier, I earned my B.Eng. degree from School of Automation Engineering at Northeast Electric Power University (NEEPU) in Jilin, China (2020). I’m always enthusiastic about computer stuff and passionate about creating exciting products during my study and research journey.

Life and Experience


In a moment, passes sorrow. That which passes will be dear. (by Alexander Pushkin)

To some extent, I didn’t perform well in the College Entrance Examination in China and was admitted into one of the so-called 双非大学 (GROUP 2 UNIVERSITIES identified in UK and Australia). I entered Northeast Electric Power University (NEEPU), ranked 216 in China Mainland, with tremendous pressure and stress.

Northeast Electric Power University Campus

Undergraduate Graduation from College (Class 2016)


If today were the last day of my life, would I want to do what I’m about to do today? (By Steve Jobs)

Truth be told, I was so lucky because I have met very lovely professors at NEEPU, and they provided me with excellent laboratories and resources. During college, I spent almost four years at laboratories, where I cooperated with postgraduates to accomplish many projects and learn new things. I genuinely appreciate their offerings and help in the laboratory with a heart full of gratitude.

Laboratory at School of Automation Engineering, NEEPU

Coffee and milk at lab

Good Friends

Meanwhile, I have made several wonderful friends. Even now, I want to thank every one of them because I cannot become who I am without them. Seeing each other, I’m getting better w.r.t. both the personality and abilities.

Good Friends at NEEPU, Peter Zhang (张世钍), Shichang Li (李世昌)

Foreign Friends at NEEPU

Summer School at UC Irvine

I’m also excited about my summer experience at University of California, Irvine. Bay Area and Los Angeles are great places to study CS and work/intern in the related field. But I was in too much of a hurry and didn’t have a chance to visit Silicon Valley, UC Berkeley, and Stanford. What a shame! However, I was fortunate to visit San Francisco (Golden Gate Bridge), Las Vegas (Grand Canyon), and West LA.

University of California Irvine Campus

Irvine, California, USA (I love USA!)

Summer Internship at Tsinghua

Quite luckily, during my junior year, I had an opportunity to attend an internship at a lab at Tsinghua University (THU) working on Natural Language Processing tasks. It’s a fantastic experience at the best university in China.

FIT Building, Tsinghua University

Summer Internship at Philips Research

Life is like a box of chocolates. You never know what you’re gonna get. (By Forrest Gump)

During my senior year, I got an internship offer from Bosch Research China. Unfortunately, due to the COVID pandemic, I didn’t take the courage to travel to Shanghai. Later on, as soon as I’m graduating, I lost a chance to join Intel Research China since I’m no longer hold a valid student status.

While the pandemic is near over in China, I received a research internship offer from Philips Research, one of the world’s leading giant companies, and it’s well-known for medical health. Taking such an internship at a Dutch multinational company is always my dream. I’m grateful for the position, and I’m thrilled to meet so many brilliant people at Philips.

Philips Research China Building

Philips Research China Campus

My Philips Working Environment

Internship at Tencent

After Philips, I applied for an intern at Tencent, a giant tech company in China. Surprisingly, I got an offer to work on the recommendation system. Tencent is absolutely a perfect place to learn technical skills and “grow” faster. I personally like Tencent. I hope one day I can go back and work at Tencent.


Tencent 22nd Culture Day