Overview
After leading the collection and verification of a 930K-frame autonomous driving dataset, built ReliableDataset — a pipeline that removes noisy labels via model ensembles and whitens the data distribution to automatically score and prune the value of training data.
Background
Low-quality datasets yield inaccurate, biased, overfit, unfair, and unreliable models while wasting compute, time, and money — so dataset quality must be checked before training.
Approach 1
Method I · Deleting noisy labels — models trained on split data cross-check each other’s labels (ensemble voting), and a Labeling Score based on IoU and confidence filters out wrong labels by threshold.
Approach 2
Method II · Whitening the data distribution — a Diversity Score from class/box-size deviation (Z-score) is combined with the Labeling Score into a Whitening Score to sort and reduce data, finding the reduction ratio with the best accuracy.
Result 1
Before/after comparison — automatically corrected and cleaned wrong object labels for cars, pedestrians, and cyclists (e.g., ‘pedestrian’ → ‘cyclist’).
Result 2
Detection and label-cleaning results on collected data (day/night, buses, signs), compared before and after.