YOLACT: Real-Time Instance Segmentation [ICCV Trailer]

Описание: GitHub: https://github.com/dbolya/yolact
Paper: https://arxiv.org/abs/1904.02689

See you at ICCV, be there or be square.

All results were computed in real-time using 1 RTX 2080 Ti. No temporal smoothing was applied (just pass each frame individually through the network). The only training data used was COCO's 2017 train set. We did not augment our data with motion blur (though that would probably be a good idea). All videos had the same hyperparameters, and if you want to produce exactly what you see here the command used for every video was:

python eval.py --trained_model=weights/yolact_base_54_800000.pth --top_k=15 --score_threshold=0.15 --display_fps --video_multiframe=8 --video=path/to/video

From left to right, the stats in the top left corner mean 1.) What FPS were the results were generated at, 2.) The current playback rate (which is capped at the source video's framerate), 3.) The extra frames waiting to be displayed since the video FPS is slower than the rate we're generating the frames (capped at ~100 to not overflow your memory).

Videos I may or may not have used:
   • Driving Downtown - New York City 4K - USA  
   • Driving Downtown - Denver 4K - Colorado USA  
   • Vehicle stock footage  
   • GoPro: Let Me Take You To The Mountain  
   • GoPro Awards: Eagle Saves Fish from Cold  
   • GoPro: Norway to Innsbruck in 5 Days | HER...  
   • GoPro: Motocross Track Day with #9 Adam Ci...  
   • GoPro: Surfing the Galapagos with Nathan F...  
   • Winter Driving in Denver, Colorado  
   • Free HD Stock Footage Collection | Adorabl...  
   • New York City 4K - Neon Nightlife Drive  
   • Driving through Downtown New York City sou...  
   • Cute, Funny and Heart-Warming Animals | Co...  
   • A Day at Mysuru Zoo - official documentary...  

Song: Two Steps From Hell by Thomas Bergersen