zxue
zxue
A new method entitled ["Uncertainty-Guided Cross-Modal Learning for Robust Multispectral Pedestrian Detection" ](https://ieeexplore.ieee.org/stamp/stamp.jsp?tp=&arnumber=9419080) is now available, whose results in KAIST pedestrian detection are better than MBNet.
waht a great dataset! because cinic-10 consists of cifar-10 and partial imgnet downsampled images, as reported in your benchmark, do you train your model from scratch or do you train...
I encounter the compilation error with the computer configuration: cuda 10.0, pytorch1.1, python 3.7. I do not know the meaning of Apex(#1564802), but i have solved my problem with this...
作者你好， 如[这里](https://github.com/bupt-ai-cz/LLVIP/issues/11#issuecomment-1107468704)所述，LLVIP数据集自发表以后被更新了一次，更新后的LLVIP数据集的行人检测精度也提高了，即：  从图中可以看出，更新后的LLVIP仅在infrared模态上的检测结果（0.67）就优于您论文中的结果（0.636）：  考虑到LLVIP被更新过一次，于是我将您的代码在新的LLVIP上重新训练了一次，batch_size=8x4gpus, image_size=1024. cfg is [here](https://github.com/DocF/multispectral-object-detection/blob/387dc3eaa25cd6e894bb67d4c46df89b518e60d2/models/transformer/yolov5l_fusion_transformerx3_llvip.yaml), hyp is [here](https://github.com/DocF/multispectral-object-detection/blob/387dc3eaa25cd6e894bb67d4c46df89b518e60d2/data/hyp.scratch.yaml). 我训练的结果为0.656，远低于LLVIP中仅使用infrared的结果。  为了实验的公平性，我已确认您的代码和LLVIP开源的代码都是YOLOv5 (version5.0)。请看LLVIP作者的[回复](https://github.com/bupt-ai-cz/LLVIP/issues/11#issuecomment-1318706384)，以及您之前的[回复](https://github.com/DocF/multispectral-object-detection/issues/24#issuecomment-1088360685). 以上实验结果令我非常困惑，请问您是否能提供在新的LLVIP数据集上的CFT的mAP呢？这对后来者与您的方法做公平对比非常重要 期待您的回复
Yezi, hi! I just watch your live on EGSR2020, you have a good presentation, and congraduate for publishing! However, I have one question, as topic，why you group normal and roughness...
Hello dear author, This is a great work and we want to cite it in our manuscript. I notice that the AP metric is obtained via a matlab script instead...