What would be a valid/substantial achievement in a Masters project (~30 pages, 4 months) in Computer Vision/Deep Learning?
Whatever satisfies the people who actually certify the master's project in question. In my department, that would be the student's advisor (and, as a formality, the department's director of graduate studies). In other departments, that would be a committee.
Different departments and even different individual faculty, even in the same subsubfield, apply different standards for what is "valid" or "substantial", just as different publication venues (and different individual reviewers), even in the same subsubfield, apply different standards for what is "publishable".
In particular, departments necessarily set standards that are realistic for the actual population of students that they admit. Baseline expectations at MIT may be wildly unrealistic at the University of Southern North Dakota at Hoople.
Different departments (and even different faculty) also have different philosophies about what the proper goal of a master's project should be. Should it be a body of publishable research? (Or more strictly, should it be a publication at a top-tier international conference like IMCL or NeurIPS, or at least a significant chunk thereof?) Or is it sufficient to produce a useful piece of software, say for an industry client, or a well-written expository survey? Or is it sufficient to attempt publishable research and settle for a useful implementation or survey?
Obviously it is not enough to just take an existing model (e.g. Mask R-CNN or similar), find/hack a dataset by yourself, add a layer or two out-of-the-box and finetune the model to it.
No, that is not obvious. In some departments, that would be considered a valid Master's project. In my opinion, it wouldn't be a very good Master's project, but my opinion only matters for students in my department.
