The model predicts PesonKeypoints even though the target is a car

[capslocknanda]

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.

[dineshreddy91]

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

[capslocknanda]

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,

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.

Thank you in advance!

[dineshreddy91]

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

[capslocknanda]

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]

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]

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

[gsscumtseu]

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, 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.

Thank you in advance!

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