Qualcomm-AI-research / InverseForm

BSD 3-Clause Clear License
147 stars 22 forks source link

What should be the InverseForm Loss Value #23

Open nitishank opened 1 month ago

nitishank commented 1 month ago

Dear Team,

What should be the values of InverseForm loss for better output. When I am calculating then some of the elemnts of homography martix even coming negatives also.

Here is the results when I am passing input and target different: Distance_coeff: tensor([[-152.0728, -168.4718, 73.1409, 24.7386], [-179.5679, -177.8112, 71.9094, 34.5937], [-188.5097, -157.5612, 88.7863, 44.3152], [-175.3606, -168.2580, 79.7636, 50.3585]], device='cuda:0') Loss: 257.5995788574219

When I am passing both input and target same to the InverseForm : Distance_coeff: tensor([[-281.3467, -338.8790, 157.5333, 135.7935], [-240.5191, -346.4449, 44.9765, 173.4442], [-334.8512, -277.7767, 217.0034, 138.2604], [-194.6233, -254.7491, 58.1725, 121.8640]], device='cuda:0') Loss: 449.66168212890625

Can you help to explain it ? Thanks!