Doubts About the Sparsity of Detectors

[noob-dqt]

Hello author, thank you for your excellent work. I have a question about the "sparsity". I noticed in the source code that in the detection head part, sparse features are transformed into dense features, specifically x_flatten = x.dense().view(batch_size, x.features.shape[1], -1) # [B, C, H*W]. This x_flatten is also used in the decoder part. Personally, I feel that if BEV feature maps are involved, does this still qualify as sparsity? I wanted to ask, which specific part does the term 'sparse detection head' refer to in the article? Does it refer to this type of detection in the DETR paradigm?