why non-local block learns a query-independent attention map in detection task

[yaoqi-zd]

trafficstars

Hi, can you explain why non-local block learns a query-independent attention map in object detection task? Since in segmentation task, both OCNet and DANet have shown that spatial attention module (same as non-local block) can learn attention maps that concentrated on pixels with the same category as the query one, rather than query-independent.

[MendelXu]

Curious too.

[lxtGH]

Curious too. It can be sparse to query why totally independent ???

[PangYunsheng8]

I have the same question, and could you offer the visualization code?

[zhujiagang]

Curious too. And should we replace the softmax activation with sigmoid? As implied by many works like Mask R-CNN, SE net, that not competing is better.

[Hxwinchina]

Curious too.

[xvjiarui]

Sorry for the late reply. One possible explanation is that non-local block in semantics segmentation is usually inserted closed to classification layer, where the feature map is highly correlated to segmentation map. However, in the detection/instance segmentation the non-local block is inserted into backbone. We would provide semantics segmentation results in the future.

[mangoyuan]

A new work is released in Disentangled Non-Local Neural Networks.