Inspiration
We wanted to build something really unique and technical. A project which uses data science not only for Deep Learning but also allow users to play with our product and expand their imagination. We chose to combine the best of both of the worlds - Technicality, and Creativity!
What it does
The ClawPatrol basically performs two functions -
- Dinosaur Image Classification - A user can upload an image of a dinosaur and the program, using CNN Image Classification, identifies the species of the dinosaur in the image with 52% accuracy. It can identify dinosaurs of 15 different species.
Now coming to the coolest part of our project! 2) Dino Cross-Breeder - A user can use this tool to SEE what an offspring of two differen`t dinosaurs of two different species might look like! The user can select dinosaur from 47 different species and just hit "create baby" button. Just as quick as 10 seconds, the program displays a realistic image of a new imaginary dinosaur which has features of both the dinosaurs.
How we built it
1) Dinosaur Image Classification - We created a CNN, doing data cleaning, preprocessing and transformation. After that, we tried our data and got a testing accuracy of 52%. We used the adam for our non-linear function to improve efficiency and outperform our previous attempts at using only softmax, or rmsprop. We ran our CNN in on our computers and colab to get the metrics we presented.
2) Dino Cross-Breeder - First, we identified what features do we need to know about the parent dinosaurs to imagine what their offspring might look like. We identified 11 such features - such as leg size, tooth size, skin texture, etc. However there were no datasets on the internet which had the data of the features we needed against the different species of dinosaurs. So, we created a dataset which contained 11 key features of 47 different species of dinosaurs. Next, we developed an algorithm which had to work with controlled randomization to combine the features of the two parent dinosaurs so that the offspring would be suitable and at the same time would also have mixed features of the parents. Next, we connected our program with Dalle-3 API where we converted the features of the offspring in form of a prompt which produces consistent and accurate images of the desired offspring.
Challenges we ran into
We also spent a lot of time creating a CSV file of 47 different species of dinosaurs with their 11 key features, because there was no such dataset in the entire internet. We also had to work a lot towards refining the prompt for dalle-3 for it to generate consistent looking images.
Accomplishments that we're proud of
We are extremely proud of battling through the beast of convolutional neural networks in order to create an image classification program. We are also proud of our work navigating through datasets and streamlit in order to create a functional website.
What we learned
We discovered the intricacies of using machine learning to scrape through data, measuring accuracy and precision. We also learned a lot about creating our own csvs and website from scratch. It was a challenging, yet formative process.
What's next for ClawPatrolClassification
We hope to utilize our program and extend our data to be able to cover a larger variety of dinosaurs and to increase the accuracy of our model
Log in or sign up for Devpost to join the conversation.