SegFix paper link about openseg.pytorch HOT 8 CLOSED

Thanks for your interest in our work and we plan to release the paper to arXiv in the future months (as it is under submission currently).

We would like to recommend you to use the SegFix offset files to improve your results on Cityscapes semantic/instance segmentation tasks.

from openseg.pytorch.

你好，请问你论文的这些参数是用的的代码中，下面的代码做出来的吗，因为你每一个模块下，比如ocr、isa、oc等代码最后面都有下面这些代码，请问使用它来比较参数和内存和时间的吗
def count_parameters(model):
return sum(p.numel() for p in model.parameters() if p.requires_grad)

avg_time = 0
avg_mem  = 0
import time
with torch.no_grad():
    for i in range(100):
        start_time = time.time()
        outputs = conv_3x3(feats)
        outputs = baseoc_infer(outputs)
        torch.cuda.synchronize()
        avg_time += (time.time() - start_time)
        avg_mem  += (torch.cuda.max_memory_allocated()-feats.element_size() * feats.nelement())

print("Average Parameters : {}".format(count_parameters(baseoc_infer)+count_parameters(conv_3x3)))
print("Average Running Time: {}".format(avg_time/100))
print("Average GPU Memory: {}".format(avg_mem/100))

from openseg.pytorch.

请问你可视化这些过程中的张量是用的下面文件夹中的代码吗

from openseg.pytorch.

大佬，我不是质疑你的工作哈，我是感觉ocr和acfnet有一点点相似

from openseg.pytorch.

@swjtulinxi ocr是2019年初的工作，最多是同时期工作。

from openseg.pytorch.

好的，其实我有一个疑惑就是他为什么就能超过ocnet，如果前端一样哈，

from openseg.pytorch.

他用ASPP做baseline了~

from openseg.pytorch.

我是说ocrnet，在这里面其实已经没有用到non-local注意力的**了，但反而精度还要高于ocnet，在acfnet中用了aspp做baseline才超过ocnet的，大佬你这里面是在前端直接加入ocr模块，就可以超过ocnet了

from openseg.pytorch.