Feature instances not learned parameters?
Closed this issue · 5 comments
Hi,
The grads are turned off for instance features. How are they updated?
We initialize instance features as all zeros and fix their gradients, preventing them from being updated during training. This approach performs slightly better than using random initialization followed by gradient updates.
Hi, The grads are turned off for instance features. How are they updated?
How are instance features then learned?
The initial layer's instance features do not need to be learnable.
Can you explain bit more? How and where are instance features encoded then?
Instance features primarily stem from the image feature map, obtaining meaningful information through deformable aggregation.
Can you explain bit more? How and where are instance features encoded then?