AAI_2025_Capstone_Chronicles_Combined

18

Draw, Detect, Navigate ​

​

​

​

​

​

​

​

​

to recognize critical route elements altogether. With the current model’s strong performance

metrics, this model demonstrates that it is suitable for effectively addressing these challenges.

The model successfully learned the features of the drawn classes it was trained upon and with

regular evaluation, retraining on additional doodle classes, and ongoing validation in realistic

scenarios will further ensure our model meets the high standards required for practical, and

impactful deployment.

Future work will focus on scaling and domain adaption by expanding the doodle class

set and increasing the size and diversity of the training and validation data. With more classes,

and the expansion of our data set to include more images for training and validation, the

project’s overall architecture could possibly be changed. The introduction of new classes could

change the model's architecture based on the doodles having very similar properties, and

features within the images themselves. Further testing will vary input resolution and analyze

confusion between similar classes (e.g., fire truck vs. van ) to assess whether increasing model

depth or modifying feature extraction improves accuracy. With deployment to specific domains,

having an intentional imbalance favoring the classes that are conventionally found within the

domain, would be objectively important for domain-specific model accuracy. Just like avoiding

classes that are not conventionally found would highly favor non-biases within the same said

domain. It would also be imperative to have more transformations and overlaps with certain

drawings that fall within reason.

In a deployment setting, the merits and tradeoffs of other implementations warrant

further exploration. While this project embeds the model within the application on the device,

an application that transmitted data to a model endpoint in the cloud would have the ability to

leverage more powerful and complex models. This would come at the expense of required

internet connectivity and increased latency as images are transmitted to the model and results

are streamed back. It would also introduce additional constraints such as current load and

server capacity. Some latency and server capacity issues could be mitigated through distributed

computing, but the need to transmit the data will inevitably reduce speed in these cases. While

45