# Continual Object Detection: A review of definitions, strategies, and challenges

**Authors:** Angelo G. Menezes, Gustavo de Moura, Cézanne Alves, André C. P. L. F. de Carvalho  
**arXiv:** [2205.15445](https://arxiv.org/abs/2205.15445)  
**Date:** 2022-05-30  
**Summary page:** https://www.summarizepaper.com/en/paper/2205.15445v1/

## Key points

- Continual Learning is a research field that focuses on learning consecutive tasks without losing performance on previously learned ones
- There is a growing need for continual object detection in robotics and autonomous vehicles
- Continual object detection involves instances of classes that are unknown at the time but can appear in subsequent tasks as new classes to be learned
- Researchers have proposed various strategies for class-incremental object detection (CIOD)
- One strategy involves knowledge distillation, where intermediate features of the base model are distilled along with detector outputs
- Chen et al. presented the first work that made use of this type of distillation through what they named a "hint loss," while Peng et al. introduced an additional adaptive distillation step on the features and RPN outputs.
- The review analyzed 26 research papers related to CIOD and provided insights into main benchmarks, metrics, and state-of-the-art performance.
- The review provides a comprehensive evaluation of existing approaches using a new metric designed to quantify stability and plasticity systematically.
- Current trends within continual object detection are identified and possible future research directions are discussed.

## Summary

Continual Learning is a field of research that focuses on the ability to learn consecutive tasks without losing performance on previously learned ones. While most research in this area has been centered around incremental classification tasks, there is a growing need for continual object detection due to its vast range of applications in robotics and autonomous vehicles. This scenario is more complex than conventional classification as it involves instances of classes that are unknown at the time but can appear in subsequent tasks as new classes to be learned. To address this challenge, researchers have proposed various strategies for class-incremental object detection (CIOD). One such strategy involves knowledge distillation, where intermediate features of the base model are distilled along with detector outputs. Chen et al. presented the first work that made use of this type of distillation through what they named a "hint loss," while Peng et al. introduced an additional adaptive distillation step on the features and RPN outputs. In this review, we analyze the current strategies proposed to tackle CIOD and provide answers to several research questions related to this field. We identified 26 research papers that met our inclusion criteria and provided insights into main benchmarks, metrics, and state-of-the-art performance. The first proposed strategy for CIOD dates back to 2017 in the seminal paper written by Shmelkov et al., which presented a solid baseline for the problem and indirectly guided comparisons across subsequent works. Overall, our review provides a comprehensive evaluation of existing approaches using a new metric designed to quantify stability and plasticity systematically. We also identify current trends within continual object detection and discuss possible future research directions.