Video inference support is live 🚀

[fcakyon]

Thanks to @madenburak, starting from v0.9.3, sahi brings full support for sliced/standard inference for video input format for all MMDetection/Detectron2/YOLOv5 models 🔥

• Just give video file as source:

sahi predict --model_path yolov5s.pt --model_type yolov5 --source video.mp4

You can also view video render during video inference with --view_video:

sahi predict --model_path yolov5s.pt --model_type yolov5 --source video.mp4 --view_video

• To forward 100 frames, on opened window press key D
• To revert 100 frames, on opened window press key A
• To forward 20 frames, on opened window press key G
• To revert 20 frames, on opened window press key F
• To exit, on opened window press key Esc

Note: If --view_video is slow, you can add --frame_skip_interval=20 argument to skip interval of 20 frames each time.

[kiki4488]

Hello, we were trying to do object detection task with video file. But we followed your command but it didn't work well. For example, we put a video file, new video was created. we checked that file, not have detected box.

When we used yolov5's video detected module, that did work well...

How to use yolov5+Sahi for video ?

[fcakyon]

@kiki4488 can you write your parameters and code used for inference?

[kiki4488]

@fcakyon

set detection model

detection_model = Yolov5DetectionModel(
config_path='../config/coco.yaml',
model_path='../model/best.pt',
confidence_threshold=0.3,
device="cuda:0", # or 'cuda:0'
)

prediction for video

predict(
detection_model = detection_model,
source='../data/test/test_2.mp4',
slice_height=256,
slice_width=256,
overlap_height_ratio=0.2,
overlap_width_ratio=0.2,
return_dict=True,
view_video=False,
frame_skip_interval=20
)

[fcakyon]

@kiki4488 what is your video size? Can you send the link for video file so that I can test it?

[kiki4488]

@fcakyon our video's width is 528 , and height is 888
Is there a problem with the size?
The video is copyrighted, I'm so sorry..

[Akansel-AKYOL]

Is it possible to get video inference with class filtering? Normally yolov5 provides this by using --classes argument in detect.py file. How can I this with sahi's predict function. for example the model detects 80 classes but I want to use 5 classes

[fcakyon]

@Akansel-AKYOL for yolov5 models you can do it as:

from sahi import AutoDetectionModel
from sahi.predict import predict
import yolov5

model = yolov5.load('yolov5s.pt')

model.conf = 0.25  # NMS confidence threshold
model.iou = 0.45  # NMS IoU threshold
model.classes = [0, 1, 5]  # perform detection on only several classes

detection_model = AutoDetectionModel.from_pretrained(
    model_type='yolov5',
    model=model,
)

predict(
    detection_model=detection_model,
    other parameters ...
)

[Akansel-AKYOL]

@fcakyon thank you

[abhavgoel]

@fcakyon how can i modify confidence_threshold for video inference? i tried changing it in predict.py, ut it doesnt seem to work

[PankajBarai]

Can we used this for real time..webcam?

[AhmedIbrahimai]

yes , you can use this code
https://github.com/AhmedIbrahimai/SAHI-and-YOLO8-in-real-time

[utility-aagrawal]

@fcakyon Can you please look into #944 and tell me what you think? Thanks!

[AhmedIbrahimai]

sahi with yolov8 in real time
https://github.com/AhmedIbrahimai/SAHI-and-YOLO8-in-real-time