uber-research
About label['roidb']['segmes']
HI, for coco dataset, when I print label['roidb']['segmes'] in resnet_upsnet.py, I found sometimes the last segmentation is not polygon format, which is of list format, but they look like RLE format, for example, a dict {"counts': ,'size':}. What's more, the counts part is not binary string but polygon list, which is really confusing.
I have no idea why this could happen, could you please help me?
I give some example:
{'counts': [20939, 13, 289, 18, 12, 1, 271, 35, 7, 12, 248, 56, 245, 59, 242, 62, 218, 1, 19, 65, 217, 2, 17, 67, 216, 4, 14, 69, 216, 4, 13, 70, 216, 5, 12, 70, 216, 6, 11, 70, 216, 6, 11, 70, 216, 7, 11, 68, 217, 8, 10, 67, 218, 8, 1 0, 65, 220, 9, 9, 56, 229, 10, 8, 57, 228, 10, 9, 57, 227, 11, 8, 61, 223, 12, 8, 63, 220, 13, 7, 64, 220, 13, 7, 64, 220, 13, 6, 67, 219, 12, 6, 67, 220, 11, 6, 67, 221, 10, 5, 68, 225, 6, 5, 25, 9, 34, 223, 8, 4, 22, 13, 33, 223, 8, 5, 16, 19, 32, 223, 8, 5, 15, 20, 32, 223, 8, 5, 2, 1, 11, 22, 31, 223, 8, 5, 2, 5, 5, 25, 30, 223, 9, 4, 2, 37, 28, 223, 9, 5, 1, 39, 25, 224, 10, 4, 1, 47, 16, 225, 15, 53 , 8, 227, 15, 288, 15, 288, 15, 288, 15, 288, 14, 289, 13, 290, 12, 290, 13, 290, 13, 290, 13, 290, 13, 290, 12, 290, 13, 290, 13, 290, 13, 290, 13, 290, 13, 289, 13, 290, 13, 289, 14, 289, 14, 288, 15, 2 88, 15, 287, 15, 288, 15, 287, 16, 287, 15, 288, 15, 287, 12, 291, 11, 292, 9, 294, 8, 295, 8, 295, 8, 295, 7, 297, 6, 298, 5, 300, 2, 5750, 2, 300, 4, 299, 4, 299, 5, 298, 5, 297, 7, 296, 7, 296, 8, 295, 8, 295, 9, 293, 10, 293, 11, 292, 11, 292, 12, 291, 12, 291, 13, 290, 13, 4, 1, 285, 14, 2, 2, 285, 14, 1, 3, 286, 17, 286, 17, 286, 17, 286, 17, 286, 17, 286, 17, 286, 17, 286, 17, 286, 17, 286, 17, 286 , 17, 286, 17, 286, 17, 286, 17, 286, 17, 286, 17, 286, 17, 286, 17, 286, 17, 286, 17, 286, 17, 286, 17, 286, 18, 285, 18, 286, 17, 287, 16, 289, 14, 290, 13, 290, 13, 291, 11, 293, 9, 296, 5, 48167, 6, 2 96, 8, 294, 10, 292, 12, 291, 12, 34, 7, 250, 12, 1, 11, 21, 9, 249, 29, 15, 11, 248, 33, 10, 13, 247, 34, 9, 13, 247, 35, 8, 13, 247, 37, 5, 14, 247, 38, 2, 16, 247, 56, 247, 56, 247, 56, 247, 56, 247, 5 6, 247, 56, 247, 56, 247, 56, 247, 56, 247, 56, 247, 56, 247, 55, 248, 55, 248, 54, 249, 54, 249, 54, 249, 14, 5, 35, 250, 12, 11, 30, 251, 10, 19, 23, 253, 7, 26, 17, 293, 10, 294, 8, 296, 6, 298, 5, 298 , 4, 299, 4, 299, 4, 299, 4, 299, 3, 300, 2, 300, 1, 26312], 'size': [303, 500]}]
[{'size': [375, 500], 'counts': b'm_18\;5L3N2M2N2O1O100000O1OL0WE[Oh:d06O1100000NKAVE>i:FUE:j::TGXOh6i0VI[Oh6f0RIAl6
0QICn62_GGa1:n6NaGJ_1:o6:QIGn69RIHm65VIKi65XIKh65XIKh64YILg65YIJg66ZIIf6b0PI]OP71eGJ\\14o62hGG\\14l65fGI_10l67dGJc1Kj6m0ZInNg6;hGMX:MnE3R:MnE4Q:KQF4o9JSF6m9HUF8l9D\\F7c:O00M3O1N2BCaE>^:=1000000O100O1O2N2N1N3N2L4N2M^_1KgN4N100K4N200O1FB[E>d:FXE;h::0O000100QF_Oa8JdFh0i0B_8f0G]O]8d0bG^O\\8c0cG@[80eGBY8>gGDW8<iGEV8;jGFU8:kGFU8:kGGT89lGGU88kGHU87YGTO5e0c86VGYO5a0f85TG]O4>h84TG@3<j82TGC1;k81TGGN9n8@dGW1]8hNcGX1]8hNbGX1_8hNGY18gNZG^1g8cNVG^1l8bNRG^1n8cNQGn00lNo8;mFi0b9f00O01O10O1O11_FQNX9Y2N1O\NhFj0W9VOjFi0V9WOkFh0U9XOkFh0V9VOkFj0V9TOlFk0U9SOlFm0V9QOjFV1n8lNQGV1l8kNTGV1j8kNVGV1h8kNXGU1g8lNZGS1e8nN\GQ1d8oN]GP1c8PO]GP1c8POWGV1i8jNUGW1l8i01O2ROUGNZ9^1fFdNZ9Z1gFfNZ9X1gFhNY9W1hFiN1Ij8W1[GPOIMj8S1]GROFMl88RG3=JBMo85SGE21;f0a8BGF0h0X8J_H67JbH5^7JcH6]7JcH6]7IeH5]7IeH5\\7JfH4[7MgHOZ73eHK\\76eHG\\7:oG]OB78<THZOf8e0kFD[9<[FZOIa0i96]F[OGc0j92]F79I_F99FF<6nNk8e0PG02mNl8b0SGe0LkNP9>VGR1i8mNXGS1h8lNYGT1g8lNYGT1g8lNYGT1g8lNYG T1g8lNYG<EIQ9LZG:IFn80YG7MGk82XGJ;2^84WGJo96RFHo98RFFo9:RFDo9<f01O1O000000000WF@[80cGD[8<dGF[8:dGH[88\\GCUO5_98ZG3f8MYG4h8KWG6i8JWG:e8F[G>a8B_G0_8@aGa0]8@cGa0\8_OdGa0\8_OeG0[8@fG?Z8AgG>Y8BjG;V8ElG9T8 GlG8U8HkG7V8HkGITOLR9;jGId86^GHd87]GGd88^GFc8:]1O11O1O10O00O1000000O1000000cF^Oe7b0RH^OjN1S9a0UH^OfN2T90iG^OXO1K1T90hGDPO12KV90hG0oNAY9?hG5X8JiG6W8JiG6W8JiG5X8KgG5Z8KdG6]8JbG6_8IaG9^8FcG:]8EdG;\8CfG=Z 8BfG?Z8@gG0Y8@gG0Y8AfG?Z8AeG0[8@cGa0^8_OaGa08_O_Ga0b8_O^G0c8@]Ga0b8^O_Gc08]OGd0_8[ObGe0^8[ObGe0^8[OcGd0]8[OeGd0[8\\OfGc0Z8]OhGa0X8_OiG?X8@kG=V8BlG<U8CmG;Y2DViMg0L3N2N100000000000O2N2N2MYS1[O\mN4M 2N2O00000O100000O0000010O1M3O10O011N4L001O10002N3RE_O_:o0N1O1N100000000000000000O100O2N2N2O7IO1O1O1O10000000000000O1O001N20000000001O1N2N3MO012N2N2N1O10000000000000000001O1N2N7J0O01O10000O101O0O1O1000K5N2 O1OGOnD1k:8TEGj:<VECi:?WE@h:a071O00001O004M1O0006KN1O1O0001O1O1O4L8G2N3MUci0'}]]
@JoyHuYY1412 Hi, I met this problem when using this code. That's because the "crowd instances" are included. You can refer to #33 .
Please let me know if you have any problem.
@JoyHuYY1412 Hi, I met this problem when using this code. That's because the "crowd instances" are included. You can refer to #33 .
Please let me know if you have any problem.
Yes, after I edited my files as https://github.com/uber-research/UPSNet/commit/ba524d5535e47ea42216dc299fa3effad4e3e661 did, this problem is solved. Unfortunately, the performance of panoptic head decreased, for example, the panoptic loss becomes larger and accuracy becomes smaller. Have you encountered this problem?
P.S. I use coco dataset instead of cityscapes.
@JoyHuYY1412 I am sorry that I can't remember exactly. But I remember that the performance seems to be worse. The pan_loss's becoming larger is normal. Because when you drop the crowd instance, the gt of this region will be wrong, which may affect the loss.
Please let me know if you have any problem.
@JoyHuYY1412 I am sorry that I can't remember exactly. But I remember that the performance seems to be worse. The pan_loss's becoming larger is normal. Because when you drop the crowd instance, the gt of this region will be wrong, which may affect the loss.
Please let me know if you have any problem.
Yes, I believe when I drop the crowd instance, the gt of the crowd part will also be "unknown", but it also makes sense when we calculate loss, right? So I was confused when the accuracy drops so much.
@JoyHuYY1412 Oh, you are right. I was wrong. It's hard to say. It seems that we need the author's help. @YuwenXiong
@JoyHuYY1412 Oh, you are right. I was wrong. It's hard to say. It seems that we need the author's help. @YuwenXiong
Thank you for your reply, but I still don't quiet understand why there will exist these kind of format in lable['roidb']['segms']. Especially the first one, it is not RLE or Polygen, ~~moreover, the code here~~ https://github.com/uber-research/UPSNet/blob/3218581a623b02a73c3334b672fc1ce0c25fdae9/upsnet/dataset/json_dataset.py#L173-L179 ~~also filtered some data, right?~~