search

dineshreddy91 / Occlusion_Net

[CVPR2019]Occlusion-Net: 2D/3D Occluded Keypoint Localization Using Graph Networks
Other
139 stars 34 forks source link

The model predicts PesonKeypoints even though the target is a car #26

Open capslocknanda opened 2 years ago

capslocknanda commented 2 years ago

Hi Dinesh, Thank you for sharing such nice work. Although while doing inference I noticed one anomaly (but maybe I am missing some settings). The model predicts PesonKeypoints even though the target is a car. Screenshot from 2022-07-18 09-14-14

dineshreddy91 commented 2 years ago

That should not be possible as the network was never trained on person keypoints. Can you explain the steps you are using.

capslocknanda commented 2 years ago

Thank you for the fast reply Dinesh! I actually solved that error but faced one more. I am using your pre-trained model occlusion_net.pth but did not get a good prediction. the prediction looks like this, demo Am I missing some scaling factor or setting in config default.py or in config yaml, please let me know. I also train one model for 220000 iterations and got this result which is still not satisfactory. demo

Thank you in advance!

dineshreddy91 commented 2 years ago

are you using the docker file provided or some other config file?

capslocknanda commented 2 years ago

Hi Dinesh, I am using occlusion_net_test.yaml this config file from your repo.

MODEL: META_ARCHITECTURE: "GeneralizedRCNN" WEIGHT: "/models/occlusion_net.pth" BACKBONE: CONV_BODY: "R-50-FPN" RESNETS: BACKBONE_OUT_CHANNELS: 256 RPN: USE_FPN: True ANCHOR_STRIDE: (4, 8, 16, 32, 64) PRE_NMS_TOP_N_TRAIN: 2000 PRE_NMS_TOP_N_TEST: 1000 POST_NMS_TOP_N_TEST: 1000 FPN_POST_NMS_TOP_N_TEST: 1000 ROI_HEADS: USE_FPN: True

ROI_BOX_HEAD: POOLER_RESOLUTION: 7 POOLER_SCALES: (0.25, 0.125, 0.0625, 0.03125) POOLER_SAMPLING_RATIO: 2 FEATURE_EXTRACTOR: "FPN2MLPFeatureExtractor" PREDICTOR: "FPNPredictor" NUM_CLASSES: 2 KEYPOINT_ON: True ROI_KEYPOINT_HEAD: POOLER_SCALES: (0.25, 0.125, 0.0625, 0.03125) FEATURE_EXTRACTOR: "KeypointRCNNFeatureExtractor" PREDICTOR: "KeypointRCNNPredictor" POOLER_RESOLUTION: 14 POOLER_SAMPLING_RATIO: 2 RESOLUTION: 56 SHARE_BOX_FEATURE_EXTRACTOR: False NUM_CLASSES: 14 GRAPH_ON: True ROI_GRAPH_HEAD: FEATURE_EXTRACTOR: "graphRCNNFeatureExtractor" SHARE_BOX_FEATURE_EXTRACTOR: False KGNN2D: True DATASETS: TRAIN: ("keypoints_carfusion_test_cocostyle", ) TEST: ("keypoints_carfusion_train_cocostyle",) INPUT: MIN_SIZE_TRAIN: (640, 672, 704, 736, 768, 800) DATALOADER: SIZE_DIVISIBILITY: 32 SOLVER: IMS_PER_BATCH: 1 BASE_LR: 0.00025 WEIGHT_DECAY: 0.0001 STEPS: (480000, 640000) MAX_ITER: 220000 CHECKPOINT_PERIOD: 5000 TEST: IMS_PER_BATCH: 1

OUTPUT_DIR: "./log"

dineshreddy91 commented 2 years ago

I am not sure why this is happening. seems like a change in the GPUs??. or may be a newer version of pytorch is causing this issue. Generally if you used the same docker file it should work well.

gsscumtseu commented 1 year ago

I also meet this bad result, and the prediction looks like this:

/home/gss/anaconda3/envs/Occlusion_Net/lib/python3.8/site-packages/apex/init.py:68: DeprecatedFeatureWarning: apex.amp is deprecated and will be removed by the end of February 2023. Use PyTorch AMP warnings.warn(msg, DeprecatedFeatureWarning) Using MLP graph encoder. Using learned graph decoder. Using MLP graph encoder. ./log/demo.jpg 2023-09-30_12-59-35 2023-09-30_12-59-35 demo

gsscumtseu commented 1 year ago

Thank you for the fast reply Dinesh! I actually solved that error but faced one more. I am using your pre-trained model occlusion_net.pth but did not get a good prediction. the prediction looks like this, demo Am I missing some scaling factor or setting in config default.py or in config yaml, please let me know. I also train one model for 220000 iterations and got this result which is still not satisfactory. demo

Thank you in advance!

I also meet this bad results, did you solve it?