Drones are an essential tool in spotting danger from a safe distance. Xailient helps drones work faster and see better by optimizing drones’ use of bandwidth. Xailient software reduced data usage by 94% on a characteristic 1080p “in the wild” test video of a drone following a vehicle.
1Xailient cuts drone data usage by 94%
We trained a Yolo3 model and Xailient Detectum Neural Network to detect cars and trucks from an overhead drone. The Yolo3 model was tested in isolation and then combined with the Detectum using the same input dataset. The Yolo 3 model was run in Google Cloud Platform and the Detectum on a Raspberry Pi 3.
Using open-source training datasets, Xailient trained a Yolo3 object detection neural network using traditional methods. This neural network provided a standard drone-AI deployment’s “Baseline” performance.
The Baseline Yolo model was run on Google Cloud Platform and fed pre-annotated test data as input to establish a control dataset of accuracy and performance. The Baseline model had an accuracy of 92% mean average precision (mAP).
Xailient Detectum Impact
The Detectum Neural Net (DNN) was trained using the same data as the Yolo3 model. The same input test data was then run through the Xailient software before being fed into the Yolo3 baseline model.
The function of the Detectum is to identify objects of interest (cars or trucks) and compress the image background or skip images entirely with no objects of interest.
A significant benefit is to avoid transmitting data through the wireless network that has no value in object detection tasks. The Xailient software reduced the total data transferred by dropping and compressing images.
Two versions of the Detectum were tested. In the first, conservative thresholds were used to ensure no change in accuracy compared to the control.
A second version of the Detectum used a more aggressive set of configurations.
The images used in the test data were all different, with no data in common. The test data is not sequential frames from a related video. This is important because videos encoded with the MPEG standards (for example, H.264) provide frame-over-frame compression that can achieve significant data savings compared to the sequence of still images that make up the frames.
To see the impact of the Detectum on a video, the Xailient configuration was then run on “in the wild” videos of overhead drone footage.
Xailient software reduced data usage by 94% on a characteristic 1080p “in the wild” test video of a drone following a vehicle.
A second version of the Detectum used a more aggressive set of configurations. In this result, the mAP reduced from 92% to 90%, and the data reduction progressed from 63% savings to 86% (a further decrease of almost 2/3rds).
In the Real-World Video Test, total size and transmitted bit rate benefits were achieved above the MPEG compression.
The Detectum Neural Networks work best when trained for the specific deployment purpose. Training data and configuration parameters can all be adjusted to fit the purpose. The results obtained in this pilot project, while dramatic, are just a starting place in how Xailient can help drone operators.
Xailient is recruiting beta customers to test the Detectum, provide feedback about deployment environments, and further explore how the efficiency gains of Xailient software can benefit them.
Since the Detectum Neural Net transmits objects of interest in full resolution, the exact savings depend on multiple factors, including the density of objects in the input. For example, a small car on a desert road has a high “background” ratio compared with a crowded highway. Similarly, the empty desert’s search and rescue or surveillance pass will have a high “empty frames” ratio. Each of these would have different ultimate bitrates, but using the Detectum, Xailient customers can ensure they are not transmitting “useless bytes.”
Ultimately, in the real-world video test of a 1080p Drone footage, Xailient achieved a 94% reduction in data usage, a 20:1 efficiency improvement.
March 14, 2022
Xailient specializes in extremely efficient low-power computer vision. Intel's OpenVINO specializes in maximizing the performance and speed of computer vision AI workloads. OpenVINO improved Xailient FPS 9.5x on Intel hardware to 448 FPS. Together, Xailient-Intel outperforms the comparable MobileNet_SSD by 80x. Even after Intel worked the OpenVINO magic on MobileNet_SSD, Xailient-OpenVINO is 14x faster.
November 29, 2021
Xailient’s Face Recognition enables high-speed edge AI processing with low-power consumption using Sony’s IMX500 – a chip so small it can fit on the tip of your finger.
We see things differently in the dynamic field of computer vision AI
You’ll get insights and resources into: