Why there is a no_obj parameter?

[guanw-pku]

Thanks for this great work!

I noticed the loss functions adopted are dependent on no_obj. Could you please explain when no_obj is true and why the loss functions are dependent on it?

[cjw2021]

I'm sorry for the late reply.
Since we use the framework of Deformable DETR, we use the focal loss to set the matching cost of the object class in matcher.py and calculate the loss.

QAHOI/models/matcher.py

Lines 47 to 54 in 7eb19c5

	if not self.no_obj:
	alpha = 0.25
	gamma = 2.0
	neg_cost_class = (1 - alpha) * (out_obj_prob ** gamma) * (-(1 - out_obj_prob + 1e-8).log())
	pos_cost_class = alpha * ((1 - out_obj_prob) ** gamma) * (-(out_obj_prob + 1e-8).log())
	cost_obj_class = pos_cost_class[:, tgt_obj_labels] - neg_cost_class[:, tgt_obj_labels]
	else:
	cost_obj_class = -out_obj_prob[:, tgt_obj_labels]

QAHOI/models/QAHOI.py

Lines 257 to 265 in 7eb19c5

	if not self.no_obj:
	target_classes_onehot = torch.zeros([src_logits.shape[0], src_logits.shape[1], src_logits.shape[2] + 1],
	dtype=src_logits.dtype, layout=src_logits.layout, device=src_logits.device)
	target_classes_onehot.scatter_(2, target_classes.unsqueeze(-1), 1)
	target_classes_onehot = target_classes_onehot[:,:,:-1]
	loss_obj_ce = sigmoid_focal_loss(src_logits, target_classes_onehot, num_interactions, alpha=self.focal_alpha, gamma=2) * src_logits.shape[1]
	else:
	loss_obj_ce = F.cross_entropy(src_logits.transpose(1, 2), target_classes, self.empty_weight)

If using the same setting as DETR, which we call the "no_obj" setting, the matcher and loss will be the same as QPIC.

Since QAHOI uses 300 queries and in HICO-DET there are more negative results, the accuracy of setting "no_obj=False" is slightly higher.