Fast-ACVNet
Fast-ACVNet copied to clipboard
gangweiX
[TPAMI 2023] Fast-ACV: Fast Attention Concatenation Volume for Accurate and Real-time Stereo Matching
请问Fast-acv-plus.py中Fast_ACVNet_plus类的forward函数中，这段代码有什么作用？ match_left = self.desc(self.conv(features_left[0])) match_right = self.desc(self.conv(features_right[0])) 其中： self.desc = nn.Conv2d(48, 48, kernel_size=1, padding=0, stride=1) 个人觉得这个操作只是做了通道数不变的1*1卷积，有什么特别的作用吗？
Hi, I'm applying the KITTI pretrained FastACV on a new dataset (KITTI360), and trying to convert the predicted disparity of stereo pair to depth value, and the process I do...
I find that gpu utilization is too low, is there any way to improve gpu utilization
https://github.com/pcb9382/StereoAlgorithms
Excellent work! But I just wonder how can I use this network to process video stream so that I can get a similar result, just as shown in your demo.
https://github.com/pcb9382/StereoAlgorithms
首先感谢您的开源代码！ 然后我发现您的Fast_ACV_plus.py这个文件里面在模型最后的两个sort和一个gather是不是可以省略掉以节省时间，代码部分如下： att_weights_prob = F.softmax(att_weights, dim=2) _, ind = att_weights_prob.sort(2, True) k = 24 ind_k = ind[:, :, :k] ind_k = ind_k.sort(2, False)[0] att_topk = torch.gather(att_weights_prob, 2, ind_k) 上面操作中先对att_weights_prob进行排序，得到下标矩阵ind，然后再对下标矩阵的一部分ind_k进行升序排列，现在ind_k里面的对应维度的数值变成了0，1，2，3，4，5，6······，再去通过gather去取att_weights_prob中的数值，如果k等于att_weights_prob的第2个维度的值，而不是24，那att_topk就等于att_weights_prob吧？如果把k的值设成att_weights的dim=2的维度值，那么这两个sort和gather是否就完全没必要存在了呢？
Hi, in the pretrained checkpoint link I saw generalization.ckpt, I wonder what's the different between it and other pretrained model on sceneflow and KITTI? thanks
尊敬的作者您好，感谢您的工作，我在复现您的代码，在复现的过程中sceneflow数据集上训练FAST-ACVNet时，出现了masks[idx].float().mean() too small的警告；请问我该如何解决这个问题，
