A Benchmark and Baseline for Open-Set Incremental Object Detection

In real-world applications, detectors are expected to evolve and improve their perceptual abilities through incremental learning of the unknown. Open world object detection (OWOD) simulates this process by recognizing unknown objects and incrementally learning these unknown classes as labeled data becomes available. However, two issues exist in the training images of OWOD datasets: instances of both 1) unknown classes and 2) previously known classes are present. To solve them, this paper proposes Open-Set Incremental Object Detection (OSIOD), which is defined as two processes: learning from the known, and an infinite cyclical process consisting of detection of the known and the unknown, together with incremental learning of the unknown. We construct a benchmark dataset by filtering out images containing unknown and/or previously known instances, which can also be used to create a new incremental object detection scenario where images with only new classes are used for incremental learning. Additionally, we propose a baseline method for OSIOD by introducing label smoothing and finetuning for unknown class discovery and catastrophic forgetting, respectively. Comprehensive experiments provide insights into the baseline. We hope that our benchmark, baseline, and insights will promote research towards OSIOD.