Skip to content
    • News
    Jan 12, 2023.

    Performance Benchmark of YOLO v5, v7 and v8

    We run YOLO v5 vs YOLO v7 vs YOLO v8 state-of-the-art object detection models head-to-head on Jetson AGX Orin and RTX 4070 Ti to find the ones with the best speed-to-accuracy balance.

    Object detection is an important and rapidly growing area of computer vision, and YOLO (You Only Look Once) is one of the most popular frameworks for object detection. YOLO v5, v7, and v8 are the latest versions of the YOLO framework, and in this blog post, we will compare their performance on the NVIDIA Jetson AGX Orin 32GB platform, the most powerful embedded AI computer, and on an RTX 4070 Ti desktop card. Keep reading to find out which version of YOLO is the best for your needs!

    NVIDIA Jetson AGX Orin and ZED stereo camera

    Performance of YOLO v5, v7 and v8 on Jetson AGX Orin

    With the release of every new YOLO, the first question we all have is, should we upgrade to the latest version? Most of the time, benchmarks are done on high-end GPUs such as A100, far from the embedded GPUs we use in production. TensoRT is rarely used, which is known to optimize most neural networks dramatically. And few benchmarks compare YOLO v7 with the Ultralytics v5 and v8 ones.

    So at Stereolabs, we’ve decided to release in 2023 a complete COCO benchmark of YOLO v5 vs. YOLO v7 vs. YOLO v8 on AGX Orin, with actual latencies using TensorRT 8.4 and JetPack 5. Here are the results.

    Here are the detailed results for all YOLOv8 vs YOLOv5 vs YOLOv7 models in 640 resolution on both NVIDIA Jetson AGX Orin (JP5) and RTX 4070 Ti (Batch 1, TRT8.4, FP16):

    MODEL

    AP

    AP0.5

    AGX ORIN (FPS)

    RTX 4070 TI (FPS)

    v5n

    28

    45.7

    370

    934

    v8n

    37.3

    52.5

    383

    1163

    v7-tiny

    37.4

    55.2

    290

    917

    v5s

    37.4

    56.8

    277

    877

    v8s

    44.9

    61.8

    260

    925

    v5m

    45.4

    64.1

    160

    586

    v8m

    50.2

    67.2

    137

    540

    v5l

    49

    67.3

    116

    446

    v7

    51.2

    69.7

    115

    452

    v8l

    52.9

    69.8

    95

    391

    v5x

    50.7

    68.9

    67

    252

    v7x

    52.9

    71.1

    77

    294

    v8x

    53.9

    71.0

    64

    236

    Some interesting findings:

    • All v8 models see a +4 to +9 mAP increase from v5 for a similar runtime.
    • v5l and v5x are now surpassed by v8m and v8l both in mAP and speed.
    • v8n is the best lightweight model in terms of accuracy and speed.
    • YOLO v8 does not provide (yet) models trained in 1280, which still makes YOLO v7 the best choice for inference at high-resolution, albeit at a lower speed.
    Using YOLO for Digital Twins

    The new YOLOv8 is a great improvement to the classic YOLOv5 object detector. A growing trend in several industries is to combine YOLO with a depth camera, such as the ZED 2i stereo camera. It allows localizing and tracking persons and objects in space for next-level awareness. Here is an example of a live digital twin creation using a Stereolabs ZED camera and the ZED SDK.

    Digital twin of an indoor scene using YOLO and ZED 2i camera

    Summary

    In conclusion, all three versions of YOLO (v5, v7 and v8) show solid performance on the Jetson Orin platform. However, based on our testing, YOLO v8 seemed to have the best performance out of the three. It improves mAP on COCO for all the variants compared to YOLO v5 while reaching similar runtimes on Orin and RTX 4070 Ti. If you’re looking for a fast and reliable object detection framework, YOLO v8 may be the right choice for you.

    You can also use the new YOLO v8 with ZED cameras to obtain 3D bounding boxes by ingesting Custom Objects into the ZED SDK.