Abstract

In this work, we develop an alternative iterative scheme where a new subset of objects is detected at each iteration. Detected boxes from previous iterations are passed to the network in subsequent iterations to ensure that the same object is not detected twice. This iterative scheme can be applied to both one-stage and two-stage object detectors with minor modifications to the training and inference processes.

Problem Description

Modern algorithms like NMS cannot handle crowded environments, treating two crowded objects as one object.

Contributions

Object detectors can be divided into two categories:

  • Two-stage detectors (R-CNN series)
  • One-stage detectors (YOLO)
    • Higher speed, accuracy close to two-stage detectors
    • Recently anchor-free algorithms have emerged

shortcomings of the original NMS algorithm:

  • The characteristic of selecting detections with high scores causes NMS to have problems in crowded environments

Efforts made:

  • Clustering problem of NMS
  • NMS reformulated as a scoring task
  • Application of Soft-NMS
  • Combination of Fitness-NMS and Soft-NMS
  • Adaptive NMS
  • R2NMS predicts both full boxes and visible boxes of objects
  • Adding recurrent LSTM layers for training sequence generation through Hungarian loss function
  • Layer that estimates Jaccard index as detection quality score, and a novel EM merging unit

The algorithm:

Iteratively performs detection:

  • First search for some simpler objects
  • In subsequent iterations, gradually increase suppression parameters and improve accuracy

Verifying Algorithm Feasibility

  • Using AdaptIS Toy V1 and Toy V2 datasets
  • Based on MMDetection framework
  • SGD optimizer with momentum 0.9, weight decay parameter 0.0001, and initial learning rate 0.02
  • Adam lr = 0.0001

Reference

[1] D. Rukhovich, K. Sofiiuk, D. Galeev, O. Barinova, and A. Konushin, “IterDet: Iterative Scheme for ObjectDetection in Crowded Environments,” 2020, [Online]. Available: http://arxiv.org/abs/2005.05708.