Learning Causal Features for Incremental Object Detection

TL;DR

This paper proposes an incremental causal object detection model by learning causal features, which can adapt to more tasks, and conducts several experiments which shows a causal feature without data-bias can make the model adapt to new tasks better.

Abstract

Object detection limits its recognizable categories during the training phase, in which it can not cover all objects of interest for users. To satisfy the practical necessity, the incremental learning ability of the detector becomes a critical factor for real-world applications. Unfortunately, neural networks unavoidably meet catastrophic forgetting problem when it is implemented on a new task. To this end, many incremental object detection models preserve the knowledge of previous tasks by replaying samples or distillation from previous models. However, they ignore an important factor that the performance of the model mostly depends on its feature. These models try to rouse the memory of the neural network with previous samples but not to prevent forgetting. To this end, in this paper, we propose an incremental causal object detection (ICOD) model by learning causal features, which can adapt to more tasks. Traditional object detection models, unavoidably depend on the data-bias or data-specific features to get the detection results, which can not adapt to the new task. When the model meets the requirements of incremental learning, the data-bias information is not beneficial to the new task, and the incremental learning may eliminate these features and lead to forgetting. To this end, our ICOD is introduced to learn the causal features, rather than the data-bias features when training the detector. Thus, when the model is implemented to a new task, the causal features of the old task can aid the incremental learning process to alleviate the catastrophic forgetting problem. We conduct our model on several experiments, which shows a causal feature without data-bias can make the model adapt to new tasks better. \keywords{Object detection, incremental learning, causal feature.

Figures (5)

• Figure 1: The motivation of the paper. For the original task (Source task), when training, both the data bias feature and the causal feature can minimize the loss function. But only the causal features have a reliable relationship to the detection results. For other tasks, only the causal features are useful. By reducing the influence of bias features and enhancing the causal feature, our model can achieve better adaptation to other tasks, such that the problem of catastrophic forgetting is alleviated.
• Figure 2: The causality graph of the detection model. (a). The traditional object detection model may establish a path from $Y$ to $F_{b}$, which leads to a backdoor path and create a spurious correlation between data bias information $B$ and ground truth $Y$. (b). In our model, we intervene $Y$ and get $\overline{Y}$ to guide the training of $F_{b}$, which sever the path from $Y$ to $F_{b}$. The model only depends on $F_{c}$ to predict the results.
• Figure 3: The overview of the ICOD. Our model is based on Faster-RCNN. On the top of the convolutional blocks, we implement the feature decomposer to get different kinds of features, including the data bias feature and causal features. After that, the original feature, data bias feature, and causal feature are fed into the detection loss for training.
• Figure 4: The t-SNE plot of the features in ICOD, where the original feature ($F$), causal feature ($F_{c}$), and bias feature ($F_{b}$) are presented. The visualization is based on the pascal voc 10+10 task. Features for the experiments are collected with the old model.
• Figure 5: The visualization of Detection Results. The ground truth is presented in green boxes, while the detection results are shown in red. Objects from both the new and old categories are presented.