Zhengyang Feng

@dqq813 目前没适配过coco格式训练，你可以看看需不需要修改dataset代码适配你自己的格式

@dqq813 建议检查一下类别数量等修改有没有造成下标溢出

> for key in coco: > if 'layer5' in key: > my_key = 'classifier.0.convs' + key.split('conv2d_list')[1] > else: > my_key = 'backbone.' + key > if my_voc[my_key].shape == coco[key].shape: >...

@Hugo-cell111 FYI, larger γ corresponds to larger differences in loss weighting. Since loss weighting is the core of the dynamic loss, hereby the use of the expression "emphasize". - γ1...

@1623021453 maybe add some debug log to verify if the model indeed has outputs

@1623021453 这个可能得具体看下是解析后处理有bug，还是模型原始结果没有。你看的只是keypoints是空的对吧。模型原始的score和控制点有么

@thesun9999 推理应该是resize做的，这个可能和数据集的bias有关，比如训练数据某个高度以上没有车道线。你可以尝试把车道线数据集可视化正常的图片做成cityscapes的size，看看是否正常

如果也不正常，那确实可能是个可调整的bug

@guillembraso Does that mean UDA (Unsupervised Domain Adaptation) can be applied with MOT17-train as target domain? Also, does this unsupervised learning still exclude MOT17-test?

@guillembraso Thanks. What about other sources of real data. I mean is mot17-train still the only allowed data even in a unsupervised manner?