Option 1 – Write a Short Literature Review and Design and Execute an Experiment Select at least two journal papers related to object detection in the context of construction (content should be as similar to excavator detection as you can find). Write a short literature review describing the papers. Focus on the strengths and weaknesses of the work’s experimental design. • What question were the authors trying to answer? • How did the authors establish the importance of the question? • Do you agree that their question was important? Why or why not? • Did they design their experiment optimally to answer the question? Your literature review should be between 250 and 500 words. Next, design and execute your own experiment related to excavator detection in images. I have provided you with an annotated excavator dataset which is available on Canvas. I suggest you start with one of the experiment types listed below, but you are not strictly limited to these. Describe your experimental design in your final submission. Also document all steps and results of your process. Provide a discussion of results.

Experiment Type 1 – Investigating the Performance Impact of a Feature Look through the images provided to you in the “Full Size Excavator Images” folder. Identify a feature in the images you believe will impact the performance of the object detector either positively or negatively. Look through the entire dataset and create a subset of the original dataset by identifying all images where this feature is visible. Depending on the nature of the feature, you will then test the feature’s influence in either a binary or scale test set. Binary test set. Create two groups, one in which the feature is present, and a second in which the feature is absent. Make sure that there are no other systematic variations between the two groups (i.e. control for confounding features). When you test your trained object detector on the two groups, identify the impact of the feature on the ultimate performance of the network. Scale test set. Create a test set of images were your feature is visible. Characterize your feature in each image on a scale. Make sure that there are no other systematic variations along the scale (i.e. control for confounding features). When you test your trained object detector on the test set, identify the impact of the feature on the ultimate performance of the network.

Experiment Type 2 – Investigating the Performance Impact of Neural Network Architecture Compare the object detection performance of YOLOv2 with a different neural network type. Examples: YOLOv2 with a different backbone CNN, YOLOv3, SSD, and R-CNN. Draw some conclusions on the relationship between performance and some characteristics of the two networks types. https://www.mathworks.com/help/vision/examples.htm…

Experiment Type 3 – Investigating the Performance Impact of Dataset Size Example 1: Once additional data has been labelled by the class, incorporate this into the training data and see how this improves performance on the original test set. Example 2: See if you can decrease the size of the training data by removing redundant images and see how the performance on the original test set changes.