•  Totally, 43 classes of signs are labeled and annotated for each image. The main objective is a multi-label classification task to classify the sign in one image.

•  GTSRB-Training contains the following structure. There is one directory for each of the 43  classes (0000 - 00043). Each directory contains the corresponding training images in one  class. The images are PPM images (RGB color). Files are numbered in two parts: XXXXX- YYYYY.ppm. The first part,XXXXX, represents the track number. All images of one class  with identical track numbers originate from one single physical traffic sign. The second part, YYYYY, is a running number within the track. The temporal order of the images is preserved.

•  GTSRB-Test contains the following structure.  There is one directory that contains all test images (12,630 images). The images are PPM images (RGB color). Files are numbered in ascending order: 00000.ppm to 12629.ppm.

•  GTSRB-Test-GT is a CSV file containing the annotation for GTSRB-Test. In the CSV file, it contains 8 kinds of annotations: Filename; Width; Height; Roi.X1; Roi.Y1; Roi.X2; Roi.Y2; ClassId. For example, 00000.ppm; 53; 54; 6; 5; 48; 49; 16 means Filename = 00000.ppm and ClassId = 16.

•  Since the scale of this dataset is big, it is not compulsory to use the entire dataset for your implementation, however, you need to specify clearly what data you use to train and evaluate.dr

R    Hint: To further improve the model performance, some keywords are provided for your own interest:  attention network, transformer, generative adversarial network, transfer learning, domain adaptation.

5   Method Design [8%]

4%  Students are requested to develop an learning-based automatic framework for this sign clas- sification task.  Its pipeline might include but not limited to data augmentation, machine  learning, deep learning, or a combination of these various types of methods. You can use any open-source libraries, such as Pytorch, scikit-learn, Keras and Tensorflow.

3%  Students are also requested to fine-tune their method, complete performance evaluations (following the evaluation.py given), and discuss the significance of their proposed approach.

1%  Some ablation studies should be performed to verify the performance gain caused by each individual component proposed in your approach and/or other attempts made, such as data augmentation techniques adopted, data preprocessing steps implemented and etc.

R    You can get some ideas about method design from the research papers uploaded on Canvas (Reference_Papers_Opt3.zip). These papers are well-selected to present a variety of method designs and levels of complexity.

6   Project Report [8%]

The report should contain introduction, methods, experimental setup, results & discussion, con- clusion, and references. It should be 2 – 4 pages (maximum 6) in two-column layout, using this template (http://www.latextemplates.com/template/wenneker-article) to follow the sci- entific style and formatted with LATEX1 . A brief guideline of the report sections is as follows:

1%  Introduction: introducing the project aim, methods and findings (suggested in 200 - 300 words; no more than 500 words).

2%  Methods: presenting the details of your method developed, including a brief description of method theories and details in design choices.

1%  Experimental setup: describing the dataset and evaluation metrics (suggested in 100 - 200 words; no more than 250 words).

2%  Results & Discussion: presenting the evaluation results of each method, including evalua- tion of main design choices, and if applicable presenting results from combining various  methods. Based on the experiment results, insightful discussion is expected to demonstrate  the corresponding analysis performed. Tables and figures are preferred to be used for result  demonstrations.

1%  Conclusion: summarising the study and findings (suggested in 100 - 120 words; no more than 150 words).

1%  References: listing literature and other references (papers and/or online resources).

7   Deliverable

Students are requested to create a zip file of all deliverables, including

• All the related source code (with your best model saved).

•  A PDF project report formatted with LATEX.

• A demo video to introduce your method and findings.

•  See Key Information section for submission restrictions.

R    Demo Video: The presentation of your method designs, findings, as well as your final results, should be recorded as a video, using any recording tools available (e.g., Zoom), and the expected duration is 3-5 mins (no more than 5 mins). It is suggested to use Microsoft PowerPoint/Google Slides along with the experiment results to present your model performance, following the same structure of the project report.

R    Bonus Marks: Different bonus marks will be distributed based on the ranking of the evaluation metrics, over all submissions made to Canvas:

•  top 10% - 4 marks

•  top 50% - 2 marks

•  top 85% - 1 mark

Please note that if any flaw or unfair comparisons found in the evaluation results, no bonus mark will be awarded and corresponding marks will be deducted.

8   Google Cloud Platform Education Grant for COMP3419 Students

You can either work locally or use the remote computing platform Google Cloud Engine (click here), which is funded by Google Cloud Platform Education Grant for COMP3419 students, to gain a better computational resources (CPU, GPU,TPU, etc).