AttentionMask: Attentive, Efficient Object Proposal Generation Focusing on Small Objects

Christian Wilms and Simone Frintrop

ACCV 2018 (accepted as oral)

We propose a novel approach for class-agnostic object proposal generation, which is efficient and especially well-suited to detect small objects. Efficiency is achieved by scale-specific objectness attention maps which focus the processing on promising parts of the image and reduce the amount of sampled windows strongly. This leads to a system, which is 33% faster than the state-of-the-art and clearly outperforming state-of-the-art in terms of average recall. Secondly, we add a module for detecting small objects, which are often missed by recent models. We show that this module improves the average recall for small objects by about 53%.

The figure shows the result of FastMask (middle) and the proposed AttentionMask (right) on small objects, highlighting one strength of AttentionMask to detect small objects very well. The filled colored contours denote found objects, while the red contours denote missed objects.

Method

Overview of the proposed AttentionMask system: An image is fed through the base net (yellow) and further downscaled to different scales (S₈ , S₁₆ , . . . ). At each scale a scale-specific objectness attention is calculated and windows (green) are sampled in the feature maps (blue boxes) at the most promising locations only to make calculations more efficient. Windows pruned due to low attention are marked in red. Each window is scored and an object mask is created. Due to the more efficient usage of the resources by introducing scale-specific objectness attention, we are able to split the base net and add a module (S₈ ) to better detect small objects like the tennis balls in this example.

Results

Qualitative results of SharpMaskZoom, FastMask and AttentionMask on the MS COCO dataset. The filled colored contours denote found objects, while not filled red contours denote missed objects.

Evaluation

Results on MS COCO with pixel-precise proposals. S, M, L denote small, medium, large objects.

Detailed analysis of AR^S over different number of proposals (left) and recall at 100 proposals for different IoU values (right).

Downloads

[Paper], [Supplementary Material], [Code], [Slides], [Poster], [Video]

Please cite our paper if you use our method.

@inproceedings{WilmsFrintropACCV2018,
 title = {{AttentionMask}: Attentive, Efficient Object Proposal Generation Focusing on Small Objects}, 
 author = {Christian Wilms and Simone Frintrop},
 booktitle = {Asian Conference on Computer Vision (ACCV)},
 year = {2018}
}