Abstract

• Detection performance tends to degrade with increasing IoU threshold
  • Overfitting during training, due to exponentially vanishing positive samples
  • Inference-time mismatch between the IoUs for which the detector is optimal and those of the input hypotheses

Introduction

• A detector optimized at a single IoU level is not necessarily optimal at other levels
• Higher quality detection requires a closer quality match between detector and the hypotheses it processes
• A detector can only have high quality if presented with high quality proposals.
• 그럼 그냥 IoU threshold 를 늘리면 성능이 증가 할까? → No
  • Distribution of hypotheses out of a proposal detector is usually heavily imbalanced towards low quality → high IoU thresholds lead to smaller number of positive training samples → Prone to Overfitting
  • Detection can be suboptimal when they are asked to work on the hypotheses of other quality levels

• Output IoU of a regressor is almost invariably better than the input IoU
• The output of a detector trained with a certain IoU threshold is a good distribution to train the detector of the next higher IoU threshold

Propose Cascade R-CNN, a multi-stage extension of R-CNN, where detector stages deeper into the cascade are sequentially more selective against close false positives.

Related Works

Iterative BBox at inference

• Single regression step is insufficient for accurate localization → apply BBox regression repeatedly
• Fixed IoU Threshold $u$ = 0.5 → degrades at high IoU BBoxes

• Distribution of the bonding box changes significantly after each iteration.

Integral Loss

• difficult to ask single classifier to perform uniformly well over all IoU levels.
• Ensemble?

• Prone to overfitting for high quality classifiers
• High quality classifiers are required to process low quality proposals at inference