opencv · jet981217 · Jul 7, 2024 · Jul 7, 2024 · Jul 7, 2024 · Jul 7, 2024
diff --git a/models/face_detection_yunet/README.md b/models/face_detection_yunet/README.md
@@ -8,13 +8,14 @@ Notes:
 - This model can detect **faces of pixels between around 10x10 to 300x300** due to the training scheme.
 - For details on training this model, please visit https://github.com/ShiqiYu/libfacedetection.train.
 - This ONNX model has fixed input shape, but OpenCV DNN infers on the exact shape of input image. See https://github.com/opencv/opencv_zoo/issues/44 for more information.
+- Quantization was done via Per Tensor method.
 
 Results of accuracy evaluation with [tools/eval](../../tools/eval).
 
 | Models      | Easy AP | Medium AP | Hard AP |
 | ----------- | ------- | --------- | ------- |
 | YuNet       | 0.8871  | 0.8710    | 0.7681  |
-| YuNet quant | 0.8838  | 0.8683    | 0.7676  |
+| YuNet quant | 0.8809  | 0.8626    | 0.7493  |
 
 \*: 'quant' stands for 'quantized'.
 

diff --git a/models/face_detection_yunet/face_detection_yunet_2023mar_int8.onnx b/models/face_detection_yunet/face_detection_yunet_2023mar_int8.onnx
diff --git a/models/face_recognition_sface/README.md b/models/face_recognition_sface/README.md
@@ -8,13 +8,14 @@ Note:
 - Model files encode MobileFaceNet instances trained on the SFace loss function, see the [SFace paper](https://arxiv.org/abs/2205.12010) for reference.
 - ONNX file conversions from [original code base](https://github.com/zhongyy/SFace) thanks to [Chengrui Wang](https://github.com/crywang).
 - (As of Sep 2021) Supporting 5-landmark warping for now, see below for details.
+- Quantization was done via Per Tensor method.
 
 Results of accuracy evaluation with [tools/eval](../../tools/eval).
 
 | Models      | Accuracy |
 | ----------- | -------- |
 | SFace       | 0.9940   |
-| SFace quant | 0.9932   |
+| SFace quant | 0.9928   |
 
 \*: 'quant' stands for 'quantized'.
 

diff --git a/models/face_recognition_sface/face_recognition_sface_2021dec_int8.onnx b/models/face_recognition_sface/face_recognition_sface_2021dec_int8.onnx
diff --git a/models/facial_expression_recognition/README.md b/models/facial_expression_recognition/README.md
@@ -7,6 +7,7 @@ Note:
 - Progressive Teacher is contributed by [Jing Jiang](https://scholar.google.com/citations?user=OCwcfAwAAAAJ&hl=zh-CN).
 -  [MobileFaceNet](https://link.springer.com/chapter/10.1007/978-3-319-97909-0_46) is used as the backbone and the model is able to classify seven basic facial expressions (angry, disgust, fearful, happy, neutral, sad, surprised).
 - [facial_expression_recognition_mobilefacenet_2022july.onnx](https://github.com/opencv/opencv_zoo/raw/master/models/facial_expression_recognition/facial_expression_recognition_mobilefacenet_2022july.onnx) is implemented thanks to [Chengrui Wang](https://github.com/crywang).
+- Quantization was done via Per Channel method.
 
 Results of accuracy evaluation on [RAF-DB](http://whdeng.cn/RAF/model1.html).
 

diff --git a/...ial_expression_recognition/facial_expression_recognition_mobilefacenet_2022july_int8.onnx b/...ial_expression_recognition/facial_expression_recognition_mobilefacenet_2022july_int8.onnx
diff --git a/models/handpose_estimation_mediapipe/README.md b/models/handpose_estimation_mediapipe/README.md
@@ -14,6 +14,7 @@ This model is converted from TFlite to ONNX using following tools:
 **Note**:
 - The int8-quantized model may produce invalid results due to a significant drop of accuracy.
 - Visit https://github.com/google/mediapipe/blob/master/docs/solutions/models.md#hands for models of larger scale.
+- Quantization was done via Per Tensor method.
 
 ## Demo
 

diff --git a/models/handpose_estimation_mediapipe/handpose_estimation_mediapipe_2023feb_int8.onnx b/models/handpose_estimation_mediapipe/handpose_estimation_mediapipe_2023feb_int8.onnx
diff --git a/models/human_segmentation_pphumanseg/README.md b/models/human_segmentation_pphumanseg/README.md
@@ -1,6 +1,6 @@
 # PPHumanSeg
 
-This model is ported from [PaddleHub](https://github.com/PaddlePaddle/PaddleHub) using [this script from OpenCV](https://github.com/opencv/opencv/blob/master/samples/dnn/dnn_model_runner/dnn_conversion/paddlepaddle/paddle_humanseg.py).
+This model is ported from [PaddleHub](https://github.com/PaddlePaddle/PaddleHub) using [this script from OpenCV](https://github.com/opencv/opencv/blob/master/samples/dnn/dnn_model_runner/dnn_conversion/paddlepaddle/paddle_humanseg.py). Quantization was done via Per Tensor method.
 
 ## Demo
 
@@ -47,7 +47,7 @@ Results of accuracy evaluation with [tools/eval](../../tools/eval).
 | Models             | Accuracy       | mIoU          |
 | ------------------ | -------------- | ------------- |
 | PPHumanSeg         | 0.9581         | 0.8996        |
-| PPHumanSeg quant   | 0.4365         | 0.2788        |
+| PPHumanSeg quant   | 0.7261         | 0.3687        |
 
 
 \*: 'quant' stands for 'quantized'.

diff --git a/models/human_segmentation_pphumanseg/human_segmentation_pphumanseg_2023mar_int8.onnx b/models/human_segmentation_pphumanseg/human_segmentation_pphumanseg_2023mar_int8.onnx
diff --git a/models/image_classification_mobilenet/README.md b/models/image_classification_mobilenet/README.md
@@ -6,12 +6,14 @@ MobileNetV2: Inverted Residuals and Linear Bottlenecks
 
 Results of accuracy evaluation with [tools/eval](../../tools/eval).
 
+Quantization was done via Per Channel method for V1 and Per Tensor for V2
+
 | Models             | Top-1 Accuracy | Top-5 Accuracy |
 | ------------------ | -------------- | -------------- |
 | MobileNet V1       | 67.64          | 87.97          |
-| MobileNet V1 quant | 55.53          | 78.74          |
+| MobileNet V1 quant | 40.50          | 53.87          |
 | MobileNet V2       | 69.44          | 89.23          |
-| MobileNet V2 quant | 68.37          | 88.56          |
+| MobileNet V2 quant | 58.10          | 87.40          |
 
 \*: 'quant' stands for 'quantized'.
 

diff --git a/models/image_classification_mobilenet/image_classification_mobilenetv1_2022apr_int8.onnx b/models/image_classification_mobilenet/image_classification_mobilenetv1_2022apr_int8.onnx
diff --git a/models/image_classification_mobilenet/image_classification_mobilenetv2_2022apr_int8.onnx b/models/image_classification_mobilenet/image_classification_mobilenetv2_2022apr_int8.onnx
diff --git a/models/license_plate_detection_yunet/README.md b/models/license_plate_detection_yunet/README.md
@@ -3,6 +3,7 @@
 This model is contributed by Dong Xu (徐栋) from [watrix.ai](watrix.ai) (银河水滴).
 
 Please note that the model is trained with Chinese license plates, so the detection results of other license plates with this model may be limited.
+Quantization was done via Per Tensor method.
 
 ## Demo
 

diff --git a/models/license_plate_detection_yunet/license_plate_detection_lpd_yunet_2023mar_int8.onnx b/models/license_plate_detection_yunet/license_plate_detection_lpd_yunet_2023mar_int8.onnx
diff --git a/models/object_detection_yolox/README.md b/models/object_detection_yolox/README.md
@@ -10,6 +10,7 @@ Key features of the YOLOX object detector
 
 Note:
 - This version of YoloX: YoloX_s
+- Quantization was done via Per Tensor method.
 
 ## Demo
 

diff --git a/models/object_detection_yolox/object_detection_yolox_2022nov_int8.onnx b/models/object_detection_yolox/object_detection_yolox_2022nov_int8.onnx
diff --git a/models/palm_detection_mediapipe/README.md b/models/palm_detection_mediapipe/README.md
@@ -9,6 +9,7 @@ SSD Anchors are generated from [GenMediaPipePalmDectionSSDAnchors](https://githu
 
 **Note**:
 - Visit https://github.com/google/mediapipe/blob/master/docs/solutions/models.md#hands for models of larger scale.
+- Quantization was done via Per Tensor method.
 
 ## Demo
 

diff --git a/models/palm_detection_mediapipe/palm_detection_mediapipe_2023feb_int8.onnx b/models/palm_detection_mediapipe/palm_detection_mediapipe_2023feb_int8.onnx
diff --git a/models/pose_estimation_mediapipe/README.md b/models/pose_estimation_mediapipe/README.md
@@ -10,6 +10,7 @@ This model is converted from TFlite to ONNX using following tools:
 
 **Note**:
 - Visit https://github.com/google/mediapipe/blob/master/docs/solutions/models.md#pose for models of larger scale.
+- Quantization was done via Per Channel method.
 ## Demo
 
 ### python

diff --git a/models/pose_estimation_mediapipe/pose_estimation_mediapipe_2023mar_int8.onnx b/models/pose_estimation_mediapipe/pose_estimation_mediapipe_2023mar_int8.onnx
diff --git a/models/text_recognition_crnn/README.md b/models/text_recognition_crnn/README.md
@@ -3,6 +3,7 @@
 [An End-to-End Trainable Neural Network for Image-based Sequence Recognition and Its Application to Scene Text Recognition](https://arxiv.org/abs/1507.05717)
 
 Results of accuracy evaluation with [tools/eval](../../tools/eval) at different text recognition datasets.
+2021 Sep English model's Quantization was done via Per Channel method.
 
 | Model name   | ICDAR03(%) | IIIT5k(%) | CUTE80(%) |
 | ------------ | ---------- | --------- | --------- |

diff --git a/models/text_recognition_crnn/text_recognition_CRNN_EN_2021sep_int8.onnx b/models/text_recognition_crnn/text_recognition_CRNN_EN_2021sep_int8.onnx