[vip-help] VIP-IPA: Advice for the Lane Detection Team

[Stevens, William Henry]

Hello,

This is William Stevens from the Lane Detection team. We are reaching out to vip-help because we would like to get some advice, guidance, and recommendations from the Professors and graduate students regarding the direction our project should take this semester. Last semester, we had discussions with Professor Zoltowski and Professor Delp about our desire to take on a more involved project goal for Spring 2024, with the intention of producing a final report that would be sufficient and acceptable for being uploaded to a website like arxiv.org.

Over the past few years, the lane detection team has researched and implemented various machine learning techniques in Python that succeed in performing some of the computer vision tasks relevant to autonomous driving software. This includes object detection, for locating and classifying cars, pedestrians, signs, etc. in dashboard images, as well as image segmentation, for highlighting and drawing the location and type of lanes and lane lines in the same images. Last semester, this work culminated into the research and implementation of the DETR architecture, which utilizes transformers to perform object detection.

Over the past week, our team has been brainstorming and discussing the direction our team should take this semester, with consideration of our goal to write a paper that is strong enough to be published online. However, we have been struggling to produce good ideas, especially since we feel that reenacting research done in other papers may not be worthy of our own publication. Nearly all the work the Lane Detection team has done in the past has been derivative research, and thus we are not very comfortable with the process of conducting our own research.

So far, we have considered looking into various DETR optimizations published in papers, such as Efficient-DETR, Anchor-DETR, etc. However, this project goal feels a bit underwhelming to us considering our aspirations. Another idea came from our past member Misha, who suggested that we begin to work on the next step in autonomous driving software, which aims to translate the computer vision information into something that would allow the algorithms to make decisions about its surroundings. For example, combining the object detection and lane segmentation models together and developing software that learns to translate these detections and segmentations into a 2D map to begin quantifying concepts like distance and speed in relation to the vehicles. However, we are also worried about this project goal because it may be too ambitious. We are struggling to find a middle ground here that would allow us to perform meaningful and unique research of our topic while still being reasonable and feasible to achieve in a single semester.

We would greatly appreciate input from any of the professors or graduate students regarding this issue. If anyone has any feedback, ideas, or suggestions for us, please let us know. Thank you.

Sincerely,
William Stevens
Vishal Urs, Karthik Selvaraj, Gabriel Torres, and Gaurish Lakhanpal
-------------- next part --------------
An HTML attachment was scrubbed...
URL: </ECN/mailman/archives/vip-help/attachments/20240117/8f97de8c/attachment.htm>