Attributes correction parameters#
Sometimes recognition algorithms, when using multiple classifiers, can determine attributes that contradict each other. For example, define a vehicle as a public transport and special transport at the same time.
In this case in order to understand which of the recognized attributes should be trusted, recognition accuracy threshold values and rules for interpreting the results of classifiers were proposed.
There are two types of thresholds: soft thresholds and hard thresholds.
Displaying results based on thresholds, setting them up and changing them are available in the CARS Analytics UI. See more details «CARS Analytics. User manual».
Soft thresholds#
Soft thresholds are used for basic filtering.
If there are conflicting attribute values and the accuracy of recognition of one of them is below the specified threshold, it should be assumed that the value of this attribute is erroneous.
Table 20 shows the values of soft thresholds for attributes.
Table 20. Soft thresholds values
| Attribute | Threshold value |
|---|---|
| Vehicle type | 0.8 |
| Vehicle model | 0.6 |
| Emergency type | 0.85 |
| Public transport type | 0.75 |
| Special transport type | 0.8 |
| Vehicle color | 0.92 |
Hard thresholds#
The use of hard thresholds is an additional filtering recommended over soft thresholds. Hard thresholds should be used in conjunction with the rules listed below for the most complete interpretation of the results.
Table 21 shows the values of hard thresholds for attributes.
Table 21. Hard thresholds values
| Attribute | Threshold value |
|---|---|
| Vehicle type | 0.9 |
| Vehicle model | 0.75 |
| Emergency type | 0.9 |
| Public transport type | 0.9 |
| Special transport type | 0.9 |
| Vehicle color | 0.92 |
The rules used to filter attributes at hard thresholds are presented in Tables 22 and 23:
Table 22. Attributes correction rules 1-4 for hard thresholds
| Condition 1 | Operator | Condition 2 | Result |
|---|---|---|---|
| Rule 1 | |||
| Attribute "Public transport type" == NOT "Other" | AND | Evaluation of recognition accuracy > hard threshold for the type of public transport | The attribute "Emergency type" == "Not emergency" (regardless of other ratings) and the attribute “Special transport type" == "Not special" (regardless of other ratings) |
| Rule 2 | |||
| Attribute "Vehicle type" == A | AND | Evaluation of recognition accuracy > hard threshold for vehicle type | The remaining attributes (that have defined thresholds) are not relevant |
| Rule 3 | |||
| Attribute "Vehicle type" == B_light | AND | Evaluation of recognition accuracy > hard threshold for vehicle type | Attribute “Special transport type" == "Not special" (regardless of other ratings) |
| Rule 4 | |||
| Attribute "Vehicle type" == B_heavy | AND | Evaluation of recognition accuracy > hard threshold for vehicle type | Acceptable values for the attribute "Special transport type" == "Not special" or "Truck", OTHERWISE "Undefined" |
Table 23. Attributes correction rule 5 for hard thresholds
| Condition 1 | Operator | Condition 2 | Operator | Condition 3 | Result |
|---|---|---|---|---|---|
| Rule 5 | |||||
| Evaluation of the recognition accuracy of the vehicle model < hard threshold for the vehicle model | AND | Estimation of vehicle type recognition accuracy < hard threshold for vehicle type | AND | Evaluation of the accuracy of recognition of the emergency type < hard threshold for the emergency type | The remaining attributes (that have defined thresholds) are not relevant |