RF-DETR vs YOLO26: Tiny Objects, HUGE Accuracy Leap! #shorts

追加: 2026-06-03

[00:00:00]want accuracy if you want something that can detect with more more accuracy small objects that can detect better in complex scenarios where it also understand the context around and RF that is definitely uh the best to go. Why? Because also it's the first model ever to reach the and break the record of 60% average precision on the Coco dataset.

[00:00:27]And now we're going to compare the benchmark. So if we see uh YOLO 26, let's compare the benchmarks between them for detection and for segmentations.

[00:00:37]And the RF that has six models, nano, small, medium, large, extra large, double extra large. If we check the accuracy what the RF that can achieve, it's quite amazing. While for this one, okay, there are two metrics. So where uh Ultralytics here is showing only 50 95.

[00:01:00]So this ensures that the 50 is when the expected position of the object and the predicted position intersect at least 50%. So if they intersect at least 50%, it means that the object is identified in the correct position. In the average precision 50, in the 50 95, they check the different accuracy till 95%.

[00:01:21]So in general, this one it requires much more precision. With the speed, so they test the speed on the Nvidia T4 with the optimized tensor RT version where where we have like just 2.3 milliseconds for the nano version of RF that and here we have 1.7.