Regression in box head is used twice. about toponet HOT 5 CLOSED

I see. We are currently utilizing the raw implementation of mmdet.
After carefully reviewing the code, I discover that the refinement process relies on the reference points from the last layer. To clarify, the first layer uses the init_reference points, while the second layer uses inter_references[0], which corresponds to the init_reference refined with reg_results[0].
Although this implementation may appear slightly misleading and redundant, it is indeed correct.

from toponet.

Hi, @wenjie710. We set with_box_refine=True in TE head (bbox_head), and with_box_refine=False in LC head (lane_head). We did not appliy any refinement to the reference points, as shown in the SGNN decoder. The reference points for each decoder layer are the same.

from toponet.

In our early experiments, we discovered that refining the single reference point with the predicted lane's bounding box center or middle point actually harms performance. This inspired us to propose lane attention with multiple reference points in the LaneSegNet paper.

from toponet.

@sephyli Thank you for your prompt response. To clarify, the TE head in question seems to have a refinement step applied to the output reference points, despite the fact that the refinement is already performed within self.transformer due to with_box_refine=True. Therefore, the additional refinement may be unnecessary.

from toponet.

OK I see. Thank you for explaining it so clearly.

from toponet.