Benchmarking deep learning models for real-time object detection on various platforms.
Test data is recorded with a GoPro in 1080p 30 FPS and is 31s long.
To clone this repo one may need to use Git LFS for the large files (models, data).
The tested models are all trained on MS COCO and evaluated on the personal recorded video.
The models and their score on the COCO evaluation set:
- SSD MobileNet V1 - 21 mAP
- SSD Inception V2 - 24 mAP
- Faster R-CNN Inception V2 - 28 mAP
- Faster R-CNN Inception ResNet V2 - 37 mAP
The models are deployed directly from the available notebooks.
Threading is utilized to limit the bottleneck from decoding and processing the video frames.
The tested platforms include:
- MacBook Pro 2015 2.7 GHz Intel Core i5 CPU
- NVIDIA GTX1060 6 GB GPU
- Raspberry Pi 3 model B
- NVIDIA Jetson TX2
The results are shown in the end of each notebook. - will get back and make a summary here later.
Furthermore the original video processed by SSD MobileNet and Faster R-CNN Inception ResNet is available as a performance comparison on youtube: https://www.youtube.com/watch?v=EY5XbRQylIg
None of the models are further optimized towards runtime inference. With a lot of work, the models for e.g. Jetson TX2 could be configured to run with TensorRT. This would effectively boost performance with at least a number of times - if not 10x.