YOLOV8 on openVino runtime Output shape error

[official-MKV]

I am building a computer-vision detection model.
I don't have a NVIDIA GPU. I use an intel based graphics card which makes inference on CPU look way better, but still after training a detection model using 'yolov8n.pt' and running the 'best.pt' model on my cpu, it is still terribly slow. I converted the 'best.pt' model to an openvino IR model to run on openvino-runtime( I read some articles that claim it boosts performance when running on intel-based cpu).
My problem is with the output shape of the IR model.

import cv2
import matplotlib.pyplot as plt
import numpy as np
from openvino.runtime import Core
from ultralytics import YOLO
ie = Core()

model = ie.read_model(model=r"\models\IR\best.xml")
compiled_model = ie.compile_model(model=model, device_name="CPU")

input_layer_ir = compiled_model.input(0)
output_layer_ir = compiled_model.output(0)


image = cv2.imread(r'\Resources\test_image.jpg')
N, C, H, W = input_layer_ir.shape
resized_image = cv2.resize(image, (W, H))
input_image = np.expand_dims(resized_image.transpose(2, 0, 1), 0)
plt.imshow(cv2.cvtColor(image, cv2.COLOR_BGR2RGB));
boxes = compiled_model([input_image])[output_layer_ir]

print(boxes.shape)

<Output: names[output0] shape[1,6,8400] type: f32>

I don't understand the output and how to go about it and it seems to be an output shape particualr to yolov8, other pretrained models don't have this shape. I have tried training using differnt yolov8 models and even tried different datasets but still endup with the same shape that has no meaning to me [1,6,8400]. How do I go about this ? What is this shape

[derek-knox]

@official-MKV This issue may help #751

As I understand it:

1. 1: batch (number of inputs where 1 is one image)
   • I'm a little fuzzy on the definition here FYI
2. 6: box + number of classes (first 4 = xywh of box, last 2 = probability of bounding box against each class idx—0 and 1 respectively)
3. 8400: number of detections
   • I believe this number is a function of the stride value (defaults to 32) and the input dimensions

If anyone has more clarity, detail, or correctness to add, please do.

[deno-source]

@official-MKV This issue may help #751

As I understand it:

1. 1: batch (number of inputs where 1 is one image)

   • I'm a little fuzzy on the definition here FYI

2. 6: box + number of classes (first 4 = xywh of box, last 2 = probability of bounding box against each class idx—0 and 1 respectively)

3. 8400: number of detections

   • I believe this number is a function of the stride value (defaults to 32) and the input dimensions

If anyone has more clarity, detail, or correctness to add, please do.

Hello, I have also encountered this problem, in tensorflow.js, how can I concretely handle [1,6,8400] and turn them into predictions?

[derek-knox]

@official-MKV My updated understanding: 8400 is not the "number of detections" but instead a dimensions size that is a function of the stride. As such this number is a function of the model input dimensions. By default, the stride is 32.

So for stride = 32 we have 640/32 which gives us the base cell size of 20. Doubling a few times (I'm guessing this step evolved experientially) gives us three total values:

• 20
• 40
• 80

And these are squared and summed to get the final dimension count (80^2 + 40^2 + 20^2 = 8400).

So a 256px input results in 256/32 = 8. So:

• 8
• 16
• 32

And these are squared and multiplied to get the final dimension count (32^2 + 16^2 + 8^2 = 1344).

---

https://openmmlab.medium.com/dive-into-yolov8-how-does-this-state-of-the-art-model-work-10f18f74bab1#:~:text=(2)%20Dimensional%20transformation,b%2C8400%2C4)%2C%20respectively.