AV dataset exploration-extended
Created on August 18|Last edited on August 19
Comment
Artifact Table
P1 Attiribute
Images with same p1 attribute appear to be similar, this should considered when splitting the dataset between training
Model's Bicycle Class Detection
Bicycles are rare and small, it maybe hard for our models to learn this class.
Person and Traffic light/sign
P1 Attribute
P2 Attribute
Focused on evaluating the distribution of detected objects based on the 'p1' and 'p2' attribute. By filtering the 'person' column with a value of 1, 'traffic light' column with a value of 0.8, and 'traffic sign' column with a value of 0.8, aimed to identify instances of person detection. Analysis revealed that a significant number of persons were detected in close proximity to both traffic lights and traffic signs.
Among the collected images, a total of 351 were classified as containing persons. Notably, when we refined the data by selecting images where the 'traffic light' score was equal to or exceeded 0.8, and the 'traffic sign' score was equal to or exceeded 0.8, the image count reduced to 225. This suggests that a substantial proportion of person detections occur in situations involving high-confidence traffic light and traffic sign recognition. This strong correlation indicates that our model is particularly effective at detecting individuals near these two distinct features on the road, suggesting its potential to enhance safety around such critical elements.
Group By Aggregations
Training Results
Run set
30
Add a comment