Multi-stage Reinforcement Learning for Object Detection

Abstract

We present a reinforcement learning approach for detecting objects within an image. Our approach performs a step-wise deformation of a bounding box with the goal of tightly framing the object. It uses a hierarchical tree-like representation of predefined region candidates, which the agent can zoom in on. This reduces the number of region candidates that must be evaluated so that the agent can afford to compute new feature maps before each step to enhance detection quality. We compare an approach that is based purely on zoom actions with one that is extended by a second refinement stage to fine-tune the bounding box after each zoom step. We also improve the fitting ability by allowing for different aspect ratios of the bounding box. Finally, we propose different reward functions to lead to a better guidance of the agent while following its search trajectories. Experiments indicate that each of these extensions leads to more correct detections. The best performing approach comprises a zoom stage and a refinement stage, uses aspect-ratio modifying actions and is trained using a combination of three different reward metrics.

BAR — A Reinforcement Learning Agent for Bounding-Box Automated Refinement

“…Table 5 shows that the initial detector resulted in a 7.9% AP@50, 32% of TP@50 and 68% of FP@50. BAR-DRL outperforms (König et al 2019) by increasing the AP@50 to 23.5% and the TP@50 to 57% versus a maximum achieved of 39% TP@50 for (König et al 2019) as reported in table 5. 191 191 191 195 197 198 199 197 198 194 195 194 197 199 198 199 199 199 FP@ 195 198 195 197 197 199 199 198 It can now be seen from figure 5 that even with initial proposals completely missing the object, the agent moves towards the target object.…”

“…The experiments were run on a GeForce GTX 1080 with 11,000 Exp. 2: Hierarchical Detector (Bueno et al 2017) Initialization The goal of this experiment is to test the performance of the agent on the public dataset PASCAL VOC and compare to (Bueno et al 2017) and (König et al 2019). To obtain initial proposals, a detector is trained as in (Bueno et al 2017) (Deng et al 2009), provides better initial proposals than previous methods as shown in figure 3 when trained on our 4 classes using 20 training images.…”

“…It then chooses the best sub-region to zoom in on, and repeats this process until reaching an accurate b-box. (König et al 2019) refine further the b-box by using 4 additional actions: left, right, up or down.…”

“…König et al 2019). BAR-CB is not used for this experiment, since instances of the aeroplane class vary greatly and both the algorithm and the HOG features cannot generalize well when the data changes in shape, texture, context, viewpoint etc Lin et al 2017…”

Real‐time CVSA decals recognition system using deep convolutional neural network architectures

“…In a multistage object detection pipeline system, an object's bounding box is localized, and its class identified, using different networks. König et al [31] present a multistage reinforcement learning approach for detecting objects within an image. The authors' approach is comprised of a zoom stage and a refinement stage, using aspect-ratio modifying actions, and is trained via a combination of three different reward metrics.…”