I. Scatter plot
These parameters are commonly used in the evaluation of object detection models. in:
Box
Indicates the number of all bounding boxes detected by the target detection model;val Box
Indicates the number of bounding boxes used to verify the performance of the detection model;Objectness
Represents the target object score, which is used to evaluate the probability of whether a target object is contained in a bounding box;val Objectness
Indicates the object score on the validation set;Classification
Indicates the category score, which is used to measure the category probability of a target object contained in a bounding box;val Classification
Indicates the category score on the validation set;Precision
Indicates the ratio of the number of correctly predicted positive samples to the number of all predicted positive samples, which is used to evaluate how many of the model's predicted objects are the real target objects;Recall
Indicates the ratio of the number of correctly predicted positive samples to the number of all real positive samples, which is used to evaluate whether the model can correctly detect the real target object;[email protected]
Indicates the average precision when IoU (intersection-over-union ratio) is equal to 0.5, [email protected] is the sum of the average precision of all target categories;[email protected]:0.95
Indicates the average accuracy under different IoU thresholds, where 0.5 and 0.95 represent the lower and upper bounds of IoU, respectively. [email protected]:0.95 can well reflect the comprehensiveness of the model performance.
II. Indicators
This is the output of the trained model, including the following metrics:
199/199
: Indicates the number of training rounds, currently 199, and the total number of rounds is also 199;12.5G
: Indicates the memory size occupied by the training model;0.02521
,0.02248
,0.005274
,0.05297
: Indicates the value of the loss function, which may be calculated on the training set or validation set;229
and640
: represent the number of samples and the total number of samples predicted correctly on the validation set, respectively;0.6042
and0.8467
: represent the accuracy and recall (Precision and Recall) on the verification set, respectively;0.8061
: refers to the average precision when IoU is equal to 0.5 ([email protected]);0.5687
: represents the average precision mAP of IoU from 0.5 to 0.95;0.03039
,0.02054
,0.006311
: refers to the average accuracy of each category on the validation set, which is usually used for the evaluation results of multi-category target detection models.
It can be seen from here that the training results of a naked yolov5s are not ideal
III. precision-recall_curve
precision-recall curve
It is a performance evaluation tool widely used in machine learning models. This two-dimensional curve is used to evaluate the performance of the model under different classification thresholds, where the horizontal axis represents the recall rate and the vertical axis represents the precision. As the classification threshold changes, the precision and recall of the classification results also change, thus forming some (precision, recall) coordinate points. The curve obtained by connecting these coordinate points is the "precision-recall curve". Generally, the closer the curve is to the red line in the upper left corner, the better the performance of the model, so it is necessary to make a trade-off between recall and accuracy, and choose an appropriate classification threshold for model evaluation.
IV. labels
labels.png
The file is the label file used to visualize the training process in the Yolov5 program. It shows all the different classes in the labeled data that is used for model training during training. labels.png
Each category's name and a unique color coding are included, which helps to quickly distinguish between different categories of objects when visualizing. This allows for better training, giving the model a better understanding of what each category represents. If no file is provided when running Yolov5 labels.png
, the program will automatically generate this file according to the number of categories in the dataset.
v. to visualize
Comparing the above two pictures carefully, you will find that there are many missed and wrong detections, which need to be optimized urgently.