stereo-transformer
stereo-transformer copied to clipboard
mli0603
about table5
hi, i don't really know how to get table5 in your essay without pre-train model provided by you. Did you train on sceneflow and finetune on kitti 2012 and 2015? And what this table in readme mean?
All pretrained models are provided: https://github.com/mli0603/stereo-transformer#pre-trained-models
The table is a reference for you to check if things on your end runs the same as mine. So if you run the fine-tuned model on the training set of KITTI 2015, you should get the result in first row. If you submit the test result on KITTI website, it should be the same as second row.
Thank you. I know you provide pretrain model. I want to know how to train a model without a pretrain model. I want to get the result in table5 from the start but not by using a pretrain model, can you help me about it?
------------------ 原始邮件 ------------------ 发件人: "mli0603/stereo-transformer" @.>; 发送时间: 2021年3月12日(星期五) 晚上11:39 @.>; @.@.>; 主题: Re: [mli0603/stereo-transformer] about table5 (#19)
All pretrained models are provided: https://github.com/mli0603/stereo-transformer#pre-trained-models
— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.
You can finetune the sceneflow trained model (the link above) using this script: https://github.com/mli0603/stereo-transformer/blob/main/scripts/kitti_finetune.sh
So, did you get your scenflow pretrianed model by "pre-train on Scene Flow for 15 epochs using a fixed learning rate of 1e-4 for feature extractor and Transformer, and 2e-4 for the context adjustment layer" in your essay and get your kitti finetune model by finetune the sceneflow pretrain model on kitti2015?
Yes. The pretrained script for Scene Flow can be found here: https://github.com/mli0603/stereo-transformer/blob/main/scripts/pretrain.sh
@minchong1998 Oh I see where your confusion comes from. I used 400 epochs but I wrote 300 in the finetuning script. This is a mistake when I transferring PyCharm configuration to a bash scripts. Sorry for the confusion and I will correct it.
hi，i use the kitti_finetnue.sh and finetune on kitti2012 and 2015. When i break down my process, i continue the training by loading the checkpoint like epoch_200_model.pth.tar as pretrian model and carry on. When testing, i run --inference in my terminal and store the disparity map. But i think the map is not good enough. Did i mistake some part? I didn't train 400 epoch but still 300 because i found px_error didn't change much during training.
------------------ 原始邮件 ------------------ 发件人: "mli0603/stereo-transformer" @.>; 发送时间: 2021年3月16日(星期二) 凌晨4:25 @.>; @.@.>; 主题: Re: [mli0603/stereo-transformer] about table5 (#19)
@minchong1998 Oh I see where your confusion comes from. I used 400 epochs but I wrote 300 in the finetuning script. This is a mistake when I transferring PyCharm configuration to a bash scripts. Sorry for the confusion and I will correct it.
— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or unsubscribe.
by the way, it is one of the disparity map i get. No.16 in kitti2015 test
------------------ 原始邮件 ------------------ 发件人: "mli0603/stereo-transformer" @.>; 发送时间: 2021年3月16日(星期二) 凌晨4:25 @.>; @.@.>; 主题: Re: [mli0603/stereo-transformer] about table5 (#19)
@minchong1998 Oh I see where your confusion comes from. I used 400 epochs but I wrote 300 in the finetuning script. This is a mistake when I transferring PyCharm configuration to a bash scripts. Sorry for the confusion and I will correct it.
— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or unsubscribe.
Hi @minchong1998, I attched my training log at epoch 250 and 300. Can you see if they match what you have?
Index 0, attn 0.2269, rr 0.8732, l1 0.2209, entropy 0.0019, epe 0.4879, iou 0.9571, px error 0.0094
Index 1, attn 0.2460, rr 0.7193, l1 0.2281, entropy 0.0023, epe 0.5196, iou 0.9555, px error 0.0087
Index 2, attn 0.3206, rr 0.9870, l1 0.3138, entropy 0.0016, epe 0.6210, iou 0.9695, px error 0.0137
Index 3, attn 0.2239, rr 0.7959, l1 0.1930, entropy 0.0012, epe 0.4762, iou 0.9797, px error 0.0052
Index 4, attn 0.3677, rr 0.8059, l1 0.3513, entropy 0.0010, epe 0.6586, iou 0.9760, px error 0.0185
Index 5, attn 0.1844, rr 0.7582, l1 0.1688, entropy 0.0007, epe 0.4503, iou 0.9836, px error 0.0035
Index 6, attn 0.2505, rr 0.8926, l1 0.2230, entropy 0.0007, epe 0.4844, iou 0.9842, px error 0.0138
Index 7, attn 0.3579, rr 0.7557, l1 0.3023, entropy 0.0054, epe 0.6157, iou 0.8632, px error 0.0204
Index 8, attn 0.3536, rr 0.8004, l1 0.2871, entropy 0.0016, epe 0.6047, iou 0.9668, px error 0.0106
Index 9, attn 0.3535, rr 0.7623, l1 0.2600, entropy 0.0096, epe 0.5291, iou 0.9349, px error 0.0135
Index 10, attn 0.3179, rr 0.7905, l1 0.2497, entropy 0.0017, epe 0.5338, iou 0.9710, px error 0.0121
Index 11, attn 0.4146, rr 0.8366, l1 0.3848, entropy 0.0014, epe 0.6967, iou 0.9699, px error 0.0216
Index 12, attn 0.3884, rr 0.7525, l1 0.3516, entropy 0.0014, epe 0.6665, iou 0.9805, px error 0.0157
Index 13, attn 0.6897, rr 1.0362, l1 0.6442, entropy 0.0034, epe 0.9709, iou 0.9780, px error 0.0344
Index 14, attn 0.5307, rr 0.7961, l1 0.4677, entropy 0.0086, epe 0.8167, iou 0.9589, px error 0.0229
Index 15, attn 0.4434, rr 0.8102, l1 0.3955, entropy 0.0039, epe 0.6849, iou 0.9234, px error 0.0203
Index 16, attn 0.4186, rr 0.8544, l1 0.3868, entropy 0.0011, epe 0.7135, iou 0.9817, px error 0.0120
Index 17, attn 0.4630, rr 0.8310, l1 0.4362, entropy 0.0027, epe 0.7465, iou 0.9148, px error 0.0131
Index 18, attn 0.4343, rr 0.7691, l1 0.3473, entropy 0.0017, epe 0.6633, iou 0.9561, px error 0.0254
Index 19, attn 0.2812, rr 0.8765, l1 0.2292, entropy 0.0015, epe 0.5071, iou 0.9614, px error 0.0124
Epoch 250, attn 7.2668, rr 16.5036, l1 6.4412, epe 0.6224, iou 0.9583, px error 0.0156
Index 0, attn 0.2362, rr 0.8956, l1 0.2238, entropy 0.0017, epe 0.4939, iou 0.9563, px error 0.0091
Index 1, attn 0.2385, rr 0.7332, l1 0.2241, entropy 0.0022, epe 0.5112, iou 0.9548, px error 0.0087
Index 2, attn 0.3287, rr 1.0038, l1 0.3148, entropy 0.0029, epe 0.6239, iou 0.9401, px error 0.0134
Index 3, attn 0.2230, rr 0.8094, l1 0.1949, entropy 0.0015, epe 0.4772, iou 0.9761, px error 0.0054
Index 4, attn 0.3925, rr 0.8287, l1 0.3676, entropy 0.0010, epe 0.6728, iou 0.9779, px error 0.0192
Index 5, attn 0.1788, rr 0.7786, l1 0.1579, entropy 0.0008, epe 0.4306, iou 0.9859, px error 0.0032
Index 6, attn 0.2771, rr 0.9102, l1 0.2464, entropy 0.0008, epe 0.5134, iou 0.9834, px error 0.0159
Index 7, attn 0.3520, rr 0.7600, l1 0.3059, entropy 0.0044, epe 0.6226, iou 0.8966, px error 0.0210
Index 8, attn 0.3359, rr 0.7950, l1 0.2678, entropy 0.0015, epe 0.5855, iou 0.9669, px error 0.0093
Index 9, attn 0.3510, rr 0.7809, l1 0.2603, entropy 0.0050, epe 0.5276, iou 0.9715, px error 0.0141
Index 10, attn 0.3101, rr 0.8050, l1 0.2434, entropy 0.0016, epe 0.5239, iou 0.9688, px error 0.0106
Index 11, attn 0.4138, rr 0.8499, l1 0.3961, entropy 0.0015, epe 0.7054, iou 0.9685, px error 0.0228
Index 12, attn 0.3851, rr 0.7561, l1 0.3553, entropy 0.0015, epe 0.6709, iou 0.9776, px error 0.0159
Index 13, attn 0.7022, rr 1.0670, l1 0.6817, entropy 0.0031, epe 1.0121, iou 0.9786, px error 0.0351
Index 14, attn 0.5549, rr 0.8250, l1 0.5101, entropy 0.0037, epe 0.8646, iou 0.9798, px error 0.0255
Index 15, attn 0.4236, rr 0.8251, l1 0.3801, entropy 0.0009, epe 0.6668, iou 0.9778, px error 0.0201
Index 16, attn 0.4183, rr 0.8586, l1 0.3894, entropy 0.0010, epe 0.7195, iou 0.9843, px error 0.0114
Index 17, attn 0.4427, rr 0.8347, l1 0.4201, entropy 0.0018, epe 0.7319, iou 0.9389, px error 0.0127
Index 18, attn 0.4811, rr 0.7763, l1 0.3942, entropy 0.0012, epe 0.7116, iou 0.9744, px error 0.0324
Index 19, attn 0.2778, rr 0.8895, l1 0.2294, entropy 0.0013, epe 0.5030, iou 0.9600, px error 0.0126
Epoch 300, attn 7.3234, rr 16.7827, l1 6.5632, epe 0.6284, iou 0.9659, px error 0.0162
hi, this is my log of 300. i found my iou is much less than yours and it keeps the value of 0.2144. Do you know the reason?
Index 0, l1_raw 0.2145, rr 0.9035, l1 0.2197, occ_be 0.0014, epe 0.4872, iou 0.1586, px error 0.0088 Index 1, l1_raw 0.2177, rr 0.7652, l1 0.2098, occ_be 0.0024, epe 0.4791, iou 0.1603, px error 0.0087 Index 2, l1_raw 0.3408, rr 1.0074, l1 0.3285, occ_be 0.0021, epe 0.6410, iou 0.1600, px error 0.0143 Index 3, l1_raw 0.2165, rr 0.8201, l1 0.1930, occ_be 0.0013, epe 0.4754, iou 0.2128, px error 0.0047 Index 4, l1_raw 0.3879, rr 0.8596, l1 0.3712, occ_be 0.0008, epe 0.6758, iou 0.1950, px error 0.0194 Index 5, l1_raw 0.1737, rr 0.8136, l1 0.1539, occ_be 0.0007, epe 0.4131, iou 0.2230, px error 0.0034 Index 6, l1_raw 0.2413, rr 0.9312, l1 0.2318, occ_be 0.0008, epe 0.4996, iou 0.1741, px error 0.0105 Index 7, l1_raw 0.3045, rr 0.7887, l1 0.2591, occ_be 0.0038, epe 0.5513, iou 0.1684, px error 0.0172 Index 8, l1_raw 0.3317, rr 0.8129, l1 0.2542, occ_be 0.0015, epe 0.5583, iou 0.2463, px error 0.0092 Index 9, l1_raw 0.3736, rr 0.7709, l1 0.2767, occ_be 0.0041, epe 0.5440, iou 0.2784, px error 0.0117 Index 10, l1_raw 0.3156, rr 0.8184, l1 0.2517, occ_be 0.0015, epe 0.5315, iou 0.2583, px error 0.0119 Index 11, l1_raw 0.4098, rr 0.8791, l1 0.3801, occ_be 0.0012, epe 0.6709, iou 0.1766, px error 0.0242 Index 12, l1_raw 0.3483, rr 0.7833, l1 0.3090, occ_be 0.0010, epe 0.5983, iou 0.2444, px error 0.0163 Index 13, l1_raw 0.7159, rr 1.0389, l1 0.6679, occ_be 0.0038, epe 0.9801, iou 0.2637, px error 0.0419 Index 14, l1_raw 0.4482, rr 0.7811, l1 0.4138, occ_be 0.0034, epe 0.7410, iou 0.2273, px error 0.0216 Index 15, l1_raw 0.4284, rr 0.8664, l1 0.3897, occ_be 0.0008, epe 0.6560, iou 0.2551, px error 0.0200 Index 16, l1_raw 0.3964, rr 0.8969, l1 0.3709, occ_be 0.0011, epe 0.6829, iou 0.2672, px error 0.0111 Index 17, l1_raw 0.4728, rr 0.8533, l1 0.4399, occ_be 0.0019, epe 0.7473, iou 0.2731, px error 0.0143 Index 18, l1_raw 0.4405, rr 0.7994, l1 0.3366, occ_be 0.0010, epe 0.6272, iou 0.2395, px error 0.0281 Index 19, l1_raw 0.2873, rr 0.9370, l1 0.2322, occ_be 0.0011, epe 0.5049, iou 0.1058, px error 0.0128 Epoch 300, epe 0.6032, iou 0.2144, px error 0.0158
------------------ 原始邮件 ------------------ 发件人: "mli0603/stereo-transformer" @.>; 发送时间: 2021年3月17日(星期三) 晚上10:05 @.>; @.@.>; 主题: Re: [mli0603/stereo-transformer] about table5 (#19)
Hi @minchong1998, I attched my training log at epoch 250 and 300. Can you see if they match what you have?
Index 0, attn 0.2269, rr 0.8732, l1 0.2209, entropy 0.0019, epe 0.4879, iou 0.9571, px error 0.0094 Index 1, attn 0.2460, rr 0.7193, l1 0.2281, entropy 0.0023, epe 0.5196, iou 0.9555, px error 0.0087 Index 2, attn 0.3206, rr 0.9870, l1 0.3138, entropy 0.0016, epe 0.6210, iou 0.9695, px error 0.0137 Index 3, attn 0.2239, rr 0.7959, l1 0.1930, entropy 0.0012, epe 0.4762, iou 0.9797, px error 0.0052 Index 4, attn 0.3677, rr 0.8059, l1 0.3513, entropy 0.0010, epe 0.6586, iou 0.9760, px error 0.0185 Index 5, attn 0.1844, rr 0.7582, l1 0.1688, entropy 0.0007, epe 0.4503, iou 0.9836, px error 0.0035 Index 6, attn 0.2505, rr 0.8926, l1 0.2230, entropy 0.0007, epe 0.4844, iou 0.9842, px error 0.0138 Index 7, attn 0.3579, rr 0.7557, l1 0.3023, entropy 0.0054, epe 0.6157, iou 0.8632, px error 0.0204 Index 8, attn 0.3536, rr 0.8004, l1 0.2871, entropy 0.0016, epe 0.6047, iou 0.9668, px error 0.0106 Index 9, attn 0.3535, rr 0.7623, l1 0.2600, entropy 0.0096, epe 0.5291, iou 0.9349, px error 0.0135 Index 10, attn 0.3179, rr 0.7905, l1 0.2497, entropy 0.0017, epe 0.5338, iou 0.9710, px error 0.0121 Index 11, attn 0.4146, rr 0.8366, l1 0.3848, entropy 0.0014, epe 0.6967, iou 0.9699, px error 0.0216 Index 12, attn 0.3884, rr 0.7525, l1 0.3516, entropy 0.0014, epe 0.6665, iou 0.9805, px error 0.0157 Index 13, attn 0.6897, rr 1.0362, l1 0.6442, entropy 0.0034, epe 0.9709, iou 0.9780, px error 0.0344 Index 14, attn 0.5307, rr 0.7961, l1 0.4677, entropy 0.0086, epe 0.8167, iou 0.9589, px error 0.0229 Index 15, attn 0.4434, rr 0.8102, l1 0.3955, entropy 0.0039, epe 0.6849, iou 0.9234, px error 0.0203 Index 16, attn 0.4186, rr 0.8544, l1 0.3868, entropy 0.0011, epe 0.7135, iou 0.9817, px error 0.0120 Index 17, attn 0.4630, rr 0.8310, l1 0.4362, entropy 0.0027, epe 0.7465, iou 0.9148, px error 0.0131 Index 18, attn 0.4343, rr 0.7691, l1 0.3473, entropy 0.0017, epe 0.6633, iou 0.9561, px error 0.0254 Index 19, attn 0.2812, rr 0.8765, l1 0.2292, entropy 0.0015, epe 0.5071, iou 0.9614, px error 0.0124 Epoch 250, attn 7.2668, rr 16.5036, l1 6.4412, epe 0.6224, iou 0.9583, px error 0.0156
Index 0, attn 0.2362, rr 0.8956, l1 0.2238, entropy 0.0017, epe 0.4939, iou 0.9563, px error 0.0091 Index 1, attn 0.2385, rr 0.7332, l1 0.2241, entropy 0.0022, epe 0.5112, iou 0.9548, px error 0.0087 Index 2, attn 0.3287, rr 1.0038, l1 0.3148, entropy 0.0029, epe 0.6239, iou 0.9401, px error 0.0134 Index 3, attn 0.2230, rr 0.8094, l1 0.1949, entropy 0.0015, epe 0.4772, iou 0.9761, px error 0.0054 Index 4, attn 0.3925, rr 0.8287, l1 0.3676, entropy 0.0010, epe 0.6728, iou 0.9779, px error 0.0192 Index 5, attn 0.1788, rr 0.7786, l1 0.1579, entropy 0.0008, epe 0.4306, iou 0.9859, px error 0.0032 Index 6, attn 0.2771, rr 0.9102, l1 0.2464, entropy 0.0008, epe 0.5134, iou 0.9834, px error 0.0159 Index 7, attn 0.3520, rr 0.7600, l1 0.3059, entropy 0.0044, epe 0.6226, iou 0.8966, px error 0.0210 Index 8, attn 0.3359, rr 0.7950, l1 0.2678, entropy 0.0015, epe 0.5855, iou 0.9669, px error 0.0093 Index 9, attn 0.3510, rr 0.7809, l1 0.2603, entropy 0.0050, epe 0.5276, iou 0.9715, px error 0.0141 Index 10, attn 0.3101, rr 0.8050, l1 0.2434, entropy 0.0016, epe 0.5239, iou 0.9688, px error 0.0106 Index 11, attn 0.4138, rr 0.8499, l1 0.3961, entropy 0.0015, epe 0.7054, iou 0.9685, px error 0.0228 Index 12, attn 0.3851, rr 0.7561, l1 0.3553, entropy 0.0015, epe 0.6709, iou 0.9776, px error 0.0159 Index 13, attn 0.7022, rr 1.0670, l1 0.6817, entropy 0.0031, epe 1.0121, iou 0.9786, px error 0.0351 Index 14, attn 0.5549, rr 0.8250, l1 0.5101, entropy 0.0037, epe 0.8646, iou 0.9798, px error 0.0255 Index 15, attn 0.4236, rr 0.8251, l1 0.3801, entropy 0.0009, epe 0.6668, iou 0.9778, px error 0.0201 Index 16, attn 0.4183, rr 0.8586, l1 0.3894, entropy 0.0010, epe 0.7195, iou 0.9843, px error 0.0114 Index 17, attn 0.4427, rr 0.8347, l1 0.4201, entropy 0.0018, epe 0.7319, iou 0.9389, px error 0.0127 Index 18, attn 0.4811, rr 0.7763, l1 0.3942, entropy 0.0012, epe 0.7116, iou 0.9744, px error 0.0324 Index 19, attn 0.2778, rr 0.8895, l1 0.2294, entropy 0.0013, epe 0.5030, iou 0.9600, px error 0.0126 Epoch 300, attn 7.3234, rr 16.7827, l1 6.5632, epe 0.6284, iou 0.9659, px error 0.0162
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub, or unsubscribe.
i found the groundtruth of occ_mask you use to calculate iou is inputs.occ_mask.(self.compute_iou(outputs['occ_pred'], inputs.occ_mask, loss, invalid_mask), which is set as input_data['occ_mask'] = np.zeros_like(disp).astype(np.bool) in kitti.py. This is also use to calculate entropy loss. I wonder why it is set to zero.
hi, i found that other methods in your table 5 is different with those in kitti official web. For example, d1all of cspn is 1.74 in the website and 1.63 in your table. By the way, you wrote "Evaluation 3 px Error on KITTI 2015." in the title of table 5 and i want to ask what the 3px on kitti 2015 mean? kitti2015 official web seems not provide 3px error.
------------------ 原始邮件 ------------------ 发件人: "mli0603/stereo-transformer" @.>; 发送时间: 2021年3月17日(星期三) 晚上10:05 @.>; @.@.>; 主题: Re: [mli0603/stereo-transformer] about table5 (#19)
Hi @minchong1998, I attched my training log at epoch 250 and 300. Can you see if they match what you have?
Index 0, attn 0.2269, rr 0.8732, l1 0.2209, entropy 0.0019, epe 0.4879, iou 0.9571, px error 0.0094 Index 1, attn 0.2460, rr 0.7193, l1 0.2281, entropy 0.0023, epe 0.5196, iou 0.9555, px error 0.0087 Index 2, attn 0.3206, rr 0.9870, l1 0.3138, entropy 0.0016, epe 0.6210, iou 0.9695, px error 0.0137 Index 3, attn 0.2239, rr 0.7959, l1 0.1930, entropy 0.0012, epe 0.4762, iou 0.9797, px error 0.0052 Index 4, attn 0.3677, rr 0.8059, l1 0.3513, entropy 0.0010, epe 0.6586, iou 0.9760, px error 0.0185 Index 5, attn 0.1844, rr 0.7582, l1 0.1688, entropy 0.0007, epe 0.4503, iou 0.9836, px error 0.0035 Index 6, attn 0.2505, rr 0.8926, l1 0.2230, entropy 0.0007, epe 0.4844, iou 0.9842, px error 0.0138 Index 7, attn 0.3579, rr 0.7557, l1 0.3023, entropy 0.0054, epe 0.6157, iou 0.8632, px error 0.0204 Index 8, attn 0.3536, rr 0.8004, l1 0.2871, entropy 0.0016, epe 0.6047, iou 0.9668, px error 0.0106 Index 9, attn 0.3535, rr 0.7623, l1 0.2600, entropy 0.0096, epe 0.5291, iou 0.9349, px error 0.0135 Index 10, attn 0.3179, rr 0.7905, l1 0.2497, entropy 0.0017, epe 0.5338, iou 0.9710, px error 0.0121 Index 11, attn 0.4146, rr 0.8366, l1 0.3848, entropy 0.0014, epe 0.6967, iou 0.9699, px error 0.0216 Index 12, attn 0.3884, rr 0.7525, l1 0.3516, entropy 0.0014, epe 0.6665, iou 0.9805, px error 0.0157 Index 13, attn 0.6897, rr 1.0362, l1 0.6442, entropy 0.0034, epe 0.9709, iou 0.9780, px error 0.0344 Index 14, attn 0.5307, rr 0.7961, l1 0.4677, entropy 0.0086, epe 0.8167, iou 0.9589, px error 0.0229 Index 15, attn 0.4434, rr 0.8102, l1 0.3955, entropy 0.0039, epe 0.6849, iou 0.9234, px error 0.0203 Index 16, attn 0.4186, rr 0.8544, l1 0.3868, entropy 0.0011, epe 0.7135, iou 0.9817, px error 0.0120 Index 17, attn 0.4630, rr 0.8310, l1 0.4362, entropy 0.0027, epe 0.7465, iou 0.9148, px error 0.0131 Index 18, attn 0.4343, rr 0.7691, l1 0.3473, entropy 0.0017, epe 0.6633, iou 0.9561, px error 0.0254 Index 19, attn 0.2812, rr 0.8765, l1 0.2292, entropy 0.0015, epe 0.5071, iou 0.9614, px error 0.0124 Epoch 250, attn 7.2668, rr 16.5036, l1 6.4412, epe 0.6224, iou 0.9583, px error 0.0156
Index 0, attn 0.2362, rr 0.8956, l1 0.2238, entropy 0.0017, epe 0.4939, iou 0.9563, px error 0.0091 Index 1, attn 0.2385, rr 0.7332, l1 0.2241, entropy 0.0022, epe 0.5112, iou 0.9548, px error 0.0087 Index 2, attn 0.3287, rr 1.0038, l1 0.3148, entropy 0.0029, epe 0.6239, iou 0.9401, px error 0.0134 Index 3, attn 0.2230, rr 0.8094, l1 0.1949, entropy 0.0015, epe 0.4772, iou 0.9761, px error 0.0054 Index 4, attn 0.3925, rr 0.8287, l1 0.3676, entropy 0.0010, epe 0.6728, iou 0.9779, px error 0.0192 Index 5, attn 0.1788, rr 0.7786, l1 0.1579, entropy 0.0008, epe 0.4306, iou 0.9859, px error 0.0032 Index 6, attn 0.2771, rr 0.9102, l1 0.2464, entropy 0.0008, epe 0.5134, iou 0.9834, px error 0.0159 Index 7, attn 0.3520, rr 0.7600, l1 0.3059, entropy 0.0044, epe 0.6226, iou 0.8966, px error 0.0210 Index 8, attn 0.3359, rr 0.7950, l1 0.2678, entropy 0.0015, epe 0.5855, iou 0.9669, px error 0.0093 Index 9, attn 0.3510, rr 0.7809, l1 0.2603, entropy 0.0050, epe 0.5276, iou 0.9715, px error 0.0141 Index 10, attn 0.3101, rr 0.8050, l1 0.2434, entropy 0.0016, epe 0.5239, iou 0.9688, px error 0.0106 Index 11, attn 0.4138, rr 0.8499, l1 0.3961, entropy 0.0015, epe 0.7054, iou 0.9685, px error 0.0228 Index 12, attn 0.3851, rr 0.7561, l1 0.3553, entropy 0.0015, epe 0.6709, iou 0.9776, px error 0.0159 Index 13, attn 0.7022, rr 1.0670, l1 0.6817, entropy 0.0031, epe 1.0121, iou 0.9786, px error 0.0351 Index 14, attn 0.5549, rr 0.8250, l1 0.5101, entropy 0.0037, epe 0.8646, iou 0.9798, px error 0.0255 Index 15, attn 0.4236, rr 0.8251, l1 0.3801, entropy 0.0009, epe 0.6668, iou 0.9778, px error 0.0201 Index 16, attn 0.4183, rr 0.8586, l1 0.3894, entropy 0.0010, epe 0.7195, iou 0.9843, px error 0.0114 Index 17, attn 0.4427, rr 0.8347, l1 0.4201, entropy 0.0018, epe 0.7319, iou 0.9389, px error 0.0127 Index 18, attn 0.4811, rr 0.7763, l1 0.3942, entropy 0.0012, epe 0.7116, iou 0.9744, px error 0.0324 Index 19, attn 0.2778, rr 0.8895, l1 0.2294, entropy 0.0013, epe 0.5030, iou 0.9600, px error 0.0126 Epoch 300, attn 7.3234, rr 16.7827, l1 6.5632, epe 0.6284, iou 0.9659, px error 0.0162
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub, or unsubscribe.
i found the groundtruth of occ_mask you use to calculate iou is inputs.occ_mask.(self.compute_iou(outputs['occ_pred'], inputs.occ_mask, loss, invalid_mask), which is set as input_data['occ_mask'] = np.zeros_like(disp).astype(np.bool) in kitti.py. This is also use to calculate entropy loss. I wonder why it is set to zero.
The actual occlusion mask is computed inside here. Please see Q&A 3 where I discussed this.
I updated the compute_iou function when releasing the code to make it run faster. Previously I used confusion_matrix from scipy which was very slow. Your IOU is lower because the non-valid pixels were taken into account (which is wrong), I have fixed it and will update soon. Your result looks close to my training log, so I think if you finish training and your result will be close to what's reported in the official benchmark.
hi, i found that other methods in your table 5 is different with those in kitti official web. For example, d1all of cspn is 1.74 in the website and 1.63 in your table. By the way, you wrote "Evaluation 3 px Error on KITTI 2015." in the title of table 5 and i want to ask what the 3px on kitti 2015 mean? kitti2015 official web seems not provide 3px error.
This is because we report 3px error in non-occluded region only. 3px Error means percentage of disparity error larger than 3px. You can use the filter on the top to see non-occluded result only.
the matrix provided by kitti2015 official is "onsiders a pixel to be correctly estimated if the disparity or flow end-point error is <3px or <5% ", it seems stricter than only 3px. What i confuse is where to get a result only consider <3px, but not <3px ot <5%. And where you get other methods result only consider <3px?
------------------ 原始邮件 ------------------ 发件人: "mli0603/stereo-transformer" @.>; 发送时间: 2021年3月19日(星期五) 凌晨2:15 @.>; @.@.>; 主题: Re: [mli0603/stereo-transformer] about table5 (#19)
i found the groundtruth of occ_mask you use to calculate iou is inputs.occ_mask.(self.compute_iou(outputs['occ_pred'], inputs.occ_mask, loss, invalid_mask), which is set as input_data['occ_mask'] = np.zeros_like(disp).astype(np.bool) in kitti.py. This is also use to calculate entropy loss. I wonder why it is set to zero.
The actual occlusion mask is computed inside here. Please see Q&A 3 where I discussed this.
I updated the compute_iou function when releasing the code to make it run faster. Previously I used confusion_matrix from scipy which was very slow. Your IOU is lower because the non-valid pixels were taken into account (which is wrong), I have fixed it and will update soon. Your result looks close to my training log, so I think if you finish training and your result will be close to what's reported in the official benchmark.
hi, i found that other methods in your table 5 is different with those in kitti official web. For example, d1all of cspn is 1.74 in the website and 1.63 in your table. By the way, you wrote "Evaluation 3 px Error on KITTI 2015." in the title of table 5 and i want to ask what the 3px on kitti 2015 mean? kitti2015 official web seems not provide 3px error.
This is because we report 3px error in non-occluded region only. 3px Error means percentage of disparity error larger than 3px. You can use the filter on the top to see non-occluded result only.
— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or unsubscribe.
Nice catch. It is worth updating the manuscript to further clarify this, which you are right that it will be stricter than just 3px error. The caption tries to state the result from KITTI benchmark, but defnitely not clear enough.
So did you get your result of table 5 in kitti benchmark or not?
------------------ 原始邮件 ------------------ 发件人: "mli0603/stereo-transformer" @.>; 发送时间: 2021年3月19日(星期五) 上午10:01 @.>; @.@.>; 主题: Re: [mli0603/stereo-transformer] about table5 (#19)
Nice catch. It is worth updating the manuscript to further clarify this, which you are right that it will be stricter than just 3px error. The caption tries to state the result from KITTI benchmark, but defnitely not clear enough.
— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or unsubscribe.
Yes ;) The arxiv/github results are outdated and I haven't synced them with the KITTI benchmark yet. The update will come soon (probably next week)
thank you, i understand, i'm looking forward it.
------------------ 原始邮件 ------------------ 发件人: "mli0603/stereo-transformer" @.>; 发送时间: 2021年3月19日(星期五) 上午10:15 @.>; @.@.>; 主题: Re: [mli0603/stereo-transformer] about table5 (#19)
Yes ;) The arxiv/github results are outdated and I haven't synced them with the KITTI benchmark yet. The update will come soon (probably next week)
— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or unsubscribe.
hi, i don't really know how to get table5 in your essay without pre-train model provided by you. Did you train on sceneflow and finetune on kitti 2012 and 2015? And what this table in readme mean?
您好，我可以加一下您的联系方式嘛，我现在也在看这篇论文，部署代码时有些不懂的地方，希望您能帮我解答一下，万分感谢。您也可以加我微信：yt152971
