Embedded Files
The pipeline for semantic segmentation using UNet-based architecture with MIT's ADE20K dataset
The pipeline for semantic segmentation using UNet-based architecture with MIT's ADE20K dataset
Introduced perception domain in a mechanical robot for real-time scene segmentation, achieving a 0.78 IoU score for the ground class in the testing data.
Introduced perception domain in a mechanical robot for real-time scene segmentation, achieving a 0.78 IoU score for the ground class in the testing data.
COBRA - Crater Observing Bio-inspired Rolling Articulator is a robot used to study, analyze, and gather data from impact craters on the moon with a Bio-inspired design of a snake to roll on any uneven terrain and change shape with its joints.
COBRA - Crater Observing Bio-inspired Rolling Articulator is a robot used to study, analyze, and gather data from impact craters on the moon with a Bio-inspired design of a snake to roll on any uneven terrain and change shape with its joints.
Collected and processed 10,000 data points, categorizing them into four classes using Label Studio,
Collected and processed 10,000 data points, categorizing them into four classes using Label Studio,
- Moving objects
- Immovable objects
- Ground
- Miscellaneous
We achieved a balanced and normalized data distribution with an 80% training data split.
We achieved a balanced and normalized data distribution with an 80% training data split.
COBRA has won the NASA Artemis Award, it was exciting to update its mechanism to navigate autonomously.
COBRA has won the NASA Artemis Award, it was exciting to update its mechanism to navigate autonomously.
By incorporating perception in the head of COBRA through the Intel RealSense camera and performing semantic segmentation of its scene.
By incorporating perception in the head of COBRA through the Intel RealSense camera and performing semantic segmentation of its scene.
We built our model on existing segmentation models, pre-trained on ImageNet1K, and fine-tuned on MIT's dataset – ADE20K.
We built our model on existing segmentation models, pre-trained on ImageNet1K, and fine-tuned on MIT's dataset – ADE20K.
Encoder - MobileNetv2-dilated
Encoder - MobileNetv2-dilated
Decoder - c1-deepsup
Decoder - c1-deepsup
Smoothing out the results optimized the accuracy.
Smoothing out the results optimized the accuracy.
MIT ADE20K dataset examples
MIT ADE20K dataset examples
We 3D printed its new head to accommodate the Jetson Nano and Intel camera.
We 3D printed its new head to accommodate the Jetson Nano and Intel camera.
With Jetson, post-training quantization was possible to reduce memory size by 2x.
With Jetson, post-training quantization was possible to reduce memory size by 2x.
The deployed vision models also showed a negligible difference in the accuracy of 0.78 IoU score for the ground class with faster inference.
The deployed vision models also showed a negligible difference in the accuracy of 0.78 IoU score for the ground class with faster inference.
Overall, presented a robust skill set encompassing data analysis, feature engineering, deep learning, and model optimization, contributing to the success and advancement of COBRA's capabilities in lunar exploration.
Overall, presented a robust skill set encompassing data analysis, feature engineering, deep learning, and model optimization, contributing to the success and advancement of COBRA's capabilities in lunar exploration.
Page updated
Google Sites
Report abuse