the prediction is not good

[lakehui]

Hi, I run the testing script to test the interHand2.6M's test-set. but the prediction of hand keypoints is very bad. the testing command is follow as python3.5 test.py --gpu 0 --test_epoch 20 --test_set test --annot_subset all Is there anything I might have done wrong?

[lakehui]

[mks0601]

This is my prediction and I exactly followed the README instruction. Could you show me your bash commands and results as well?

[lakehui]

the bash commands: python3.5 test.py --gpu 0 --test_epoch 20 --test_set test --annot_subset all the model is downloaded from the [InterHand2.6M v0.0]] link, and I chose InterHand2.6M_all model to test.

the changed code as follow:

1. modify vis=True in dataset.py
2. modify self.img_path & self.annot_path to myself path.
3. add code self.datalist = self.datalist[::10000] to interval select sample because testing whole test-set spend too much time.
4. modily test_batch_size = 8 in config.py the rest of code remain unchanged.

I think these changed shouldn't impact the prediction. but ... some results as follows:

[mks0601]

Hmm this is pretty weird.. Could you test again with models trained on full IH2.6M? I'll test again using models trained on IH2.6M v0.0

[lakehui]

get similar result by models of full IH2.6M

[mks0601]

The result seems like the code did not load a model and just use randomly initialized weights. Could you give me the whole bash message and double check you put the pre-trained model to the right place (output/model_dump)?

[lakehui]

`>>> Using GPU: 0 09-16 15:01:57 Creating test dataset... Load annotation from /media/data_3t/data/hand_related/interhand2.6M/InterHand2.6M/annotations/all loading annotations into memory... Done (t=8.09s) creating index... index created! Get bbox and root depth from ../data/InterHand2.6M/rootnet_output/rootnet_interhand2.6m_output_all_test.json Number of annotations in single hand sequences: 197992 Number of annotations in interacting hand sequences: 154905 09-16 15:02:31 Load checkpoint from /home/huhui/github_demo/InterHand2.6M/main/../output/model_dump/snapshot_20.pth.tar 09-16 15:02:31 Creating graph... 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 5/5 [00:02<00:00, 1.79it/s]

Evaluation start... /home/huhui/.local/lib/python3.5/site-packages/matplotlib/pyplot.py:514: RuntimeWarning: More than 20 figures have been opened. Figures created through the pyplot interface (matplotlib.pyplot.figure) are retained until explicitly closed and may consume too much memory. (To control this warning, see the rcParam figure.max_open_warning). max_open_warning, RuntimeWarning) Handedness accuracy: 0.21875 MRRPE: 170.0613021850586

MPJPE for each joint: r_thumb4: 137.19, r_thumb3: 129.61, r_thumb2: 89.98, r_thumb1: 52.77, r_index4: 181.15, r_index3: 168.56, r_index2: 132.38, r_index1: 102.33, r_middle4: 181.62, r_middle3: 166.76, r_middle2: 145.15, r_middle1: 114.99, r_ring4: 179.53, r_ring3: 173.68, r_ring2: 152.13, r_ring1: 114.38, r_pinky4: 172.50, r_pinky3: 157.58, r_pinky2: 143.62, r_pinky1: 111.21, r_wrist: 0.00, l_thumb4: 126.22, l_thumb3: 96.72, l_thumb2: 94.01, l_thumb1: 64.37, l_index4: 180.41, l_index3: 149.07, l_index2: 133.51, l_index1: 97.97, l_middle4: 187.62, l_middle3: 159.35, l_middle2: 133.43, l_middle1: 103.37, l_ring4: 166.88, l_ring3: 143.26, l_ring2: 126.90, l_ring1: 89.67, l_pinky4: 138.39, l_pinky3: 139.20, l_pinky2: 142.93, l_pinky1: 80.11, l_wrist: 0.00, MPJPE for all hand sequences: 127.63

MPJPE for each joint: r_thumb4: 136.87, r_thumb3: 120.25, r_thumb2: 74.02, r_thumb1: 43.33, r_index4: 164.27, r_index3: 141.45, r_index2: 117.04, r_index1: 86.77, r_middle4: 183.47, r_middle3: 158.92, r_middle2: 135.18, r_middle1: 101.21, r_ring4: 178.50, r_ring3: 157.69, r_ring2: 139.53, r_ring1: 99.74, r_pinky4: 162.85, r_pinky3: 141.69, r_pinky2: 138.82, r_pinky1: 94.52, r_wrist: 0.00, l_thumb4: 114.92, l_thumb3: 94.34, l_thumb2: 89.27, l_thumb1: 68.11, l_index4: 159.72, l_index3: 133.43, l_index2: 130.42, l_index1: 98.75, l_middle4: 183.97, l_middle3: 156.91, l_middle2: 127.92, l_middle1: 84.19, l_ring4: 146.85, l_ring3: 147.83, l_ring2: 115.27, l_ring1: 89.57, l_pinky4: 134.48, l_pinky3: 123.14, l_pinky2: 116.63, l_pinky1: 87.17, l_wrist: 0.00, MPJPE for single hand sequences: 118.55

MPJPE for each joint: r_thumb4: 137.44, r_thumb3: 136.81, r_thumb2: 102.26, r_thumb1: 61.35, r_index4: 194.14, r_index3: 189.42, r_index2: 144.17, r_index1: 114.30, r_middle4: 179.93, r_middle3: 172.79, r_middle2: 152.82, r_middle1: 125.59, r_ring4: 180.57, r_ring3: 185.98, r_ring2: 161.82, r_ring1: 125.64, r_pinky4: 181.27, r_pinky3: 169.80, r_pinky2: 147.31, r_pinky1: 124.05, r_wrist: 0.00, l_thumb4: 134.92, l_thumb3: 98.55, l_thumb2: 97.66, l_thumb1: 59.68, l_index4: 194.74, l_index3: 161.09, l_index2: 135.88, l_index1: 97.37, l_middle4: 190.61, l_middle3: 161.19, l_middle2: 137.66, l_middle1: 118.12, l_ring4: 184.90, l_ring3: 140.10, l_ring2: 134.95, l_ring1: 89.75, l_pinky4: 140.80, l_pinky3: 150.32, l_pinky2: 161.15, l_pinky1: 74.69, l_wrist: 0.00, MPJPE for interacting hand sequences: 134.56`

[mks0601]

Could you set trans_test = 'gt' # gt, rootnet at config.py and test again?

[lakehui]

emm.. it doesn't work.

[mks0601]

Could you download the annotation files and codes again? There is no problem on my side :(

[lakehui]

I just download these codes today. and I have check the annotation files, it is also right. I am trying to train a model.

[kingsman0000]

`>>> Using GPU: 0 09-16 15:01:57 Creating test dataset... Load annotation from /media/data_3t/data/hand_related/interhand2.6M/InterHand2.6M/annotations/all loading annotations into memory... Done (t=8.09s) creating index... index created! Get bbox and root depth from ../data/InterHand2.6M/rootnet_output/rootnet_interhand2.6m_output_all_test.json Number of annotations in single hand sequences: 197992 Number of annotations in interacting hand sequences: 154905 09-16 15:02:31 Load checkpoint from /home/huhui/github_demo/InterHand2.6M/main/../output/model_dump/snapshot_20.pth.tar 09-16 15:02:31 Creating graph... 100%|█████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 5/5 [00:02<00:00, 1.79it/s]

Evaluation start... . . .

1.How can you let the number under Creating graph...only 5?I exactly followed the README instruction. And follow is my bash message....(I'm 5759)It take me a lot of time to run test.py every time

(Hand2.6M) kingsman@kingsman-lab:~/InterHand2.6M-master_new/main$ python test.py --gpu 0 --test_epoch 20 --test_set val --annot_subset machine_annot

Using GPU: 0 09-16 20:05:27 Creating val dataset... Load annotation from ../data/InterHand2.6M/annotations/machine_annot loading annotations into memory... Done (t=3.78s) creating index... index created! Get bbox and root depth from groundtruth annotation Number of annotations in single hand sequences: 113370 Number of annotations in interacting hand sequences: 70917 09-16 20:05:53 Load checkpoint from /home/kingsman/InterHand2.6M-master_new/main/../output/model_dump/snapshot_20.pth.tar 09-16 20:05:53 Creating graph... 100%|███████████████████████████████████████| 5759/5759 [14:19<00:00, 6.70it/s]

Evaluation start...

2.After set vis=True I only see the result for single hand .What should I do for visualize two hand?Following is my output

[mks0601]

At here, you can select the test hand sequence. Currently, datalist consists of single hand and interacting hand images.

If you want to test only on interacting hand images, do self.datalist = self.datalist_ih.

[lakehui]

emm.. I have solved my bug. I finally find it is my torchvision problem. I maybe modified it at a long time ago.

[mks0601]

good for you!