Yiru SHEN
Yiru SHEN
> @Ichiruchan hello! Detecting small objects in a large image, especially when they're close to each other, can indeed be challenging. One potential solution is to adjust the `imgsz` parameter...
在ubuntu上使用了3.1.2-13107 + LiteLoaderQQNT 0.5.3 qq就无法启动,还原main值可以启动
> 在ubuntu上使用了3.1.2-13107 + LiteLoaderQQNT 0.5.3 qq就无法启动,还原main值可以启动 已解决,权限问题,致歉
我将图片缓存了下来然后立即读取,在完整的合作项目中会发生越界,而存下来的图在单独的ai推理项目中正常执行且整套流程应该是没有并行的租借读写显存的
查到问题所在了,我这里由于每次推理的图片大小不一致,yolo对象是多次构造析构的,相当于yolo.reset()并没有用,导致AtomicAdd加在了一个随机值上,(虽然我不明白为什么第一次cudaMalloc的时候没有发生这种问题?)
目前我是自己重新写了一下坐标转换,目前和python的resize结果最多差3个像素值
> 目前我是自己重新写了一下坐标转换,目前和python的resize结果最多差3个像素值 更新,看了opencv的源码改写完后只有边界部分有1个像素值的误差
And on same datasets, for semantic segmentation, i have a good result. But actually i want a instance segmentation model.
> You can try running with `retina_masks=True` Thank you very much. It works! If I want to optimize the performance of the model's segmentation part, how can I do that?...
> You can try increasing the `imgsz`. > > Or the last two values here: > > https://github.com/ultralytics/ultralytics/blob/ba949830bfd232dccd3aa531f6e1d92583ad4dbd/ultralytics/cfg/models/v8/yolov8-seg.yaml#L46 32 represents the number of channels in the feature maps, so I...