Figure 2 - uploaded by Adam Kortylewski
Content may be subject to copyright.
Download
View publication
Object detection under occlusion with RPNs and proposed robust bounding box voting. Blue box: ground truth; red box: Faster R-CNN (RPN+VGG); yellow box: RPN+CompositionalNet; green box: context-aware CompositionalNet with robust bounding box voting. Note how the RPN-based approaches fail to localize the object, while our proposed approach can accurately localize the object.

Object detection under occlusion with RPNs and proposed robust bounding box voting. Blue box: ground truth; red box: Faster R-CNN (RPN+VGG); yellow box: RPN+CompositionalNet; green box: context-aware CompositionalNet with robust bounding box voting. Note how the RPN-based approaches fail to localize the object, while our proposed approach can accurately localize the object.

Source publication
Figure 1: Bicycle detection result for an image of the MS-COCO dataset....
Figure 2: Object detection under occlusion with RPNs and proposed...
Figure 3: Example of robust bounding box voting results. Blue box:...
Figure 4: Influence of context in aeroplane detection under occlusion....
+1 Figure 5: Context segmentation results. A standard CompositionalNet...
Robust Object Detection under Occlusion with Context-Aware CompositionalNets
Preprint
Full-text available
  • May 2020
Detecting partially occluded objects is a difficult task. Our experimental results show that deep learning approaches, such as Faster R-CNN, are not robust at object detection under occlusion. Compositional convolutional neural networks (CompositionalNets) have been shown to be robust at classifying occluded objects by explicitly representing the o...
Cite
Download full-text

Contexts in source publication

Context 1
... our experiments in Section 4.1 show that RPNs cannot reliably localize strongly occluded objects. Figure 2 illustrates this limitation by depicting the detection results of Faster R-CNN trained with CutOut [8] (red box) and a combination of RPN+CompositionalNet (yellow box). We propose to address this limitation by introducing a robust part-based voting mechanism to predict the bounding box of an object based on the visible object parts (green box). ...
Context 2
... our experiments in Section 4.1 show that RPNs cannot reliably localize strongly occluded objects. Figure 2 illustrates this limitation by depicting the detection results of Faster R-CNN trained with CutOut [8] (red box) and a combination of RPN+CompositionalNet (yellow box). We propose to address this limitation by introducing a robust part-based voting mechanism to predict the bounding box of an object based on the visible object parts (green box). ...

Similar publications

Figure 2: Our model pipeline. We first regress Local Canonical...
Figure 3: The Local Canonical Coordinate (LCC) space. For each object...
Figure 5: Canonical Voting Process. For each point, we estimate its...
Figure 6: LCC Checking with Back Projection Module. Once the 3D vote...
Figure 8: Qualitative results of 3D oriented bounding box detection in...
Canonical Voting: Towards Robust Oriented Bounding Box Detection in 3D Scenes
Preprint
Full-text available
  • Nov 2020
3D object detection has attracted much attention thanks to the advances in sensors and deep learning methods for point clouds. Current state-of-the-art methods like VoteNet regress direct offset towards object centers and box orientations with an additional Multi-Layer-Perceptron network. Both their offset and orientation predictions are not accura...
View
View
Results on MOT17 test set with the private detections. ByteTrack and...
Results on MOT20 test set with the private detections. ByteTrack and...
Results on MOT17 test set with the public detections.
Results on MOT20 test set with the public detections.
+2 Results on DanceTrack test set. Methods in the bottom block use the...
Observation-Centric SORT: Rethinking SORT for Robust Multi-Object Tracking
Preprint
Full-text available
  • Mar 2022
Multi-Object Tracking (MOT) has rapidly progressed with the development of object detection and re-identification. However, motion modeling, which facilitates object association by forecasting short-term trajectories with past observations, has been relatively under-explored in recent years. Current motion models in MOT typically assume that the ob...
View