Created by: ash12358

一开始使用configs/obj365/cascade_rcnn_cls_aware_r200_vd_fpn_dcnv2_nonlocal_softnms.yml这个配置文件进行训练，但是显存不够，就把里面TrainReader部分的target_size改为[416, 448, 480, 512, 544, 576, 608, 640, 672, 704, 736, 768, 800, 832, 864, 896]，max_size改为1000，然后就可以训练了。 但是使用第73万代的权重进行测试时，却发现图片里一个目标也检测不出来，仔细研究发现不是检测不出来，而是检测出来的box框的score值太低了，所以没进行显示。 我的测试命令是： CUDA_VISIBLE_DEVICES=2 python tools/infer.py -c configs/obj365/cascade_rcnn_cls_aware_r200_vd_fpn_dcnv2_nonlocal_softnms.yml -o weights=output/cascade_rcnn_cls_aware_r200_vd_fpn_dcnv2_nonlocal_softnms/730000 --infer_img=obj365_train_000000561556.jpg

下面打印的是在这张图片里检测出的box框的socre值，从大到小排的序。

[0.19871799647808075, 0.12369368970394135, 0.10165795683860779, 0.09234470129013062, 0.07265584170818329, 0.06411097943782806, 0.062122464179992676, 0.05287851393222809, 0.04176606982946396, 0.025868283584713936, 0.022973747923970222, 0.022617541253566742, 0.015318585559725761, 0.013942703604698181, 0.011371687054634094, 0.00941875297576189, 0.008538085967302322, 0.008139370940625668, 0.008114771917462349, 0.007070728112012148, 0.0062351832166314125, 0.006038640160113573, 0.005461093503981829, 0.0053210086189210415, 0.005245765671133995, 0.004473457578569651, 0.004348668269813061, 0.004306811839342117, 0.004248030483722687, 0.003995387814939022, 0.003923956770449877, 0.003736711572855711, 0.003573616035282612, 0.0034804109018296003, 0.002887747483327985, 0.0027637837920337915, 0.0027543106116354465, 0.002701493678614497, 0.0025193614419549704, 0.0024635721929371357, 0.002446141093969345, 0.002293485216796398, 0.0022904907818883657, 0.0022407814394682646, 0.0022349494975060225, 0.0022058396134525537, 0.002164688892662525, 0.0020860133226960897, 0.0019273929065093398, 0.001894476474262774, 0.0016658041859045625, 0.0016063376097008586, 0.0015928368084132671, 0.0015067075146362185, 0.0015065361512824893, 0.0014710043324157596, 0.0014601039001718163, 0.0014205003390088677, 0.0014046811265870929, 0.001360807684250176, 0.00135509530082345, 0.0012938143918290734, 0.001271952292881906, 0.0012468107743188739, 0.00122644507791847, 0.001204382861033082, 0.0011982853757217526, 0.001098997425287962, 0.001094385632313788, 0.001076515531167388, 0.0010564541444182396, 0.0010551458690315485, 0.0010409322567284107, 0.0010178019292652607]

显卡是GeForce RTX 2080 Ti，11G。 请大家帮忙看看，现在这个情况跟修改的target_size和max_size有关吗，怎么能提高检测效果呢？