CHAMPION

Team Name: 10000SGDMRT

Winners of the TIL-AI 2023 Advanced Tier. They integrated various CV and NLP models to their robots to perform autonomous tasks such as navigation and identification of targets.

WHAT MADE THIS PROJECT UNIQUE?

1 Passing the qualifying stage with flying colours

During the qualifying stage, teams had to create and train various AI models to accomplish two (2) tasks. The first task was to correctly identify an object from a given image using object detection (OD) and re-identification (REID) models, while the second task was to translate spoken speech to text using automatic speech recognition (ASR) models.

Team 10000SGDMRT was able to develop models that were among the most consistent and accurate, which placed them top in the qualifiers.

2 Conquering the technical challenges in the finals

In the finals, teams had to integrate their OD, REID and ASR models with a robotics platform to perform autonomous navigation and solve tasks.

Team 10000SGDMRT proved that they were the best, clearing the group stage with ease and beating all other challengers to eventually winning the competition.

3 Team’s thought process in acing the challenges

The strategy to tackle the heavy tasks and tight timeline in TIL-AI was to use off-the-shelf models and pre/post processing techniques as much as possible (i.e., git clone). Paperswithcode was a website that provided us with leaderboards of SOTA models on popular tasks, which helped us in the TIL-AI challenges. Choosing the right leaderboard in Paperswithcode was also important to ensure the relevancy to the data used in TIL-AI.

DETAILS ABOUT THE PROJECT

3

Week of qualifiers

2

Days of finals

$12k

Cash prize

5

checkpoints to clear in finals

4

AI models developed

1.0

mAP@0.5 for CV performance

0.99

mAP@0.5-0.95 for CV performance

1%

WER for ASR performance

17:18

Time to complete all 5 checkpoints

TIPS FROM THE WINNING TEAM

When in doubt, just train two (2) models and ensemble. You’ll get free performance boost 99% of the time. Pay extra attention to the deployability of your solutions, especially on potential CUDA conflicts, VRAM requirements or dependency conflicts. Most of them can be verified and patched at home.

Do you think you can do better in BrainHack 2024?