Paper reading

Title

Diagnosing Rarity in Human-object Interaction Detection

Author

Mert Kilickaya, Arnold Smeulders

单位

QUvA Deep Vision Lab Amsterdam, Netherlands

论文地址 paper

https://openaccess.thecvf.com/content_CVPRW_2020/papers/w54/Kilickaya_Diagnosing_Rarity_in_Human-Object_Interaction_Detection_CVPRW_2020_paper.pdf
https://arxiv.org/abs/2006.05728

摘要 Abstract

Human-object interaction (HOI) detection is a core task in computer vision. The goal is to localize all human-object pairs and recognize their interactions. An interaction defined by a <verb, noun> tuple leads to a long-tailed visual recognition challenge since many combinations are rarely represented. The performance of the proposed models is limited especially for the tail categories, but little has been done to understand the reason. To that end, in this paper, we propose to diagnose rarity in HOI detection. We propose a three-step strategy, namely Detection, Identification and Recognition where we carefully analyse the limiting factors by studying state-of-the-art models. Our findings indicate that detection and identification steps are altered by the interaction signals like occlusion and relative location, as a result limiting the recognition accuracy.

贡献 contribution

1. propose to diagnose rarity in HOI detection
2. propose a three-step strategy, namely Detection, Identification and Recognition where we carefully analyse the limiting factors by studying state-of-the-art models.
3. findings indicate that detection and identification steps are altered by the interaction signals like occlusion and relative location, as a result limiting the recognition accuracy.

dataset

HICO-DET
The dataset has in total 47k number of images, with more than 150k human-object pair annotations.
There exists 600 distinct interaction types of which 168 are rare, for 80 unique nouns and 117 unique
verbs.

method

HO-RCNN
iCAN
TIN

网络结构 framework

image.png
image.png

性能 performance

image.png
image.png
image.png
image.png

学习体会（好的句子等）

1. Localizing human-objects of rare interactions is not challenging, however, detection is altered by the small and occluded human-objects.
2. Identifying the rare HOIs is challenging and is altered by the background clutter and human-object distance.
3. Recognizing rare interactions is influenced by the detection and identification errors, leaving a big room for improvement.
4. 本文，是cvpr2020的workshop，总结了HICO-det上的三个主要方法，做出了总结，并且指明了将来的发展方向，但是本文的贡献较小，可以作为入门学习。