The detail about the FCPose

[JosonChan]

Thanks for your nice works! But I have some questions about the implement of fcpose.

1. Does it use 2x downsampling and 8x downsampling feature to compute the mse loss together?
   

   AdelaiDet/adet/modeling/fcpose/basis_module.py

   Lines 110 to 116 in c0b2092

    if self.training and self.loss_on: 
        x = torch.cat([x, p3_logits], dim = 1) 
        x = self.upsampler(x) 
        p1_logits = self.p1_logits(x) 
        p1_loss,p3_loss = compute_loss(p1_heatmap_list, p3_heatmap_list, p1_logits, p3_logits) 
        losses['p1_loss'] = p1_loss * self.heatmap_loss_weight 
        losses['p3_loss'] = p3_loss * self.heatmap_loss_weight 

   
   2.Does it use 2x downsampling feature to compute the softmax loss? As I notice that you use the upsampler to get 2x downsampling features.
   

   AdelaiDet/adet/modeling/fcpose/fcpose_head.py

   Lines 213 to 216 in c0b2092

    mask_logits = subnetworks_forward(offsets, weights, biases, n_inst).squeeze() 
     
    mask_logits = mask_logits.reshape(-1, 17, H, W) 
    larger_mask_logits = self.upsampler(mask_logits) 

   
   

   AdelaiDet/adet/modeling/fcpose/fcpose_head.py

   Lines 230 to 233 in c0b2092

    keypoint_loss, direction_loss = \ 
        compute_loss_softmax(gt_bitmasks, larger_mask_logits, 
                        num_loss, num_instance, direction, mask_logits, gt_keypoint,  
                        max_ranges, self.distance_norm) 

Looking for your reply! Thanks

[WeianMao]

yes, supervise on larger feat can boost the performance