MidsummerHi
commented
2 months ago Thank you for your patience and assistance. While comparing FQ R-CNN, I encountered a new issue. I trained both FQ R-CNN and featenhancer+FQ R-CNN on an exdark dataset with different partitioning of training and test sets from https://github.com/cuiziteng/ICCV_MAET, using the same solver configuration. However, I found that the accuracy of the two models is very close. The configuration details are as follows:
SOLVER: AMP: ENABLED: false BACKBONE_MULTIPLIER: 1.0 BASE_LR: 2.5e-06 BASE_LR_END: 0.0 BIAS_LR_FACTOR: 1.0 CHECKPOINT_PERIOD: 5000 CLIP_GRADIENTS: CLIP_TYPE: full_model CLIP_VALUE: 1.0 ENABLED: true NORM_TYPE: 2.0 GAMMA: 0.1 # batch size IMS_PER_BATCH: 1 LR_SCHEDULER_NAME: WarmupMultiStepLR MAX_ITER: 50000 MOMENTUM: 0.9 NESTEROV: false NUM_DECAYS: 3 OPTIMIZER: ADAMW REFERENCE_WORLD_SIZE: 0 RESCALE_INTERVAL: false STEPS: - 42000 - 47000 WARMUP_FACTOR: 0.01 WARMUP_ITERS: 1000 WARMUP_METHOD: linear WEIGHT_DECAY: 0.0001 WEIGHT_DECAY_BIAS: null WEIGHT_DECAY_NORM: 0.0I think the featenhaner part is contain in
so i test FQ R-CNN in different project cloned from https://github.com/hustvl/Featurized-QueryRCNN. but they shared a similar config file when training(i kept Featurized-QueryRCNN's MODEL part)
result:
with featenhaner {"bbox/AP": 52.53054561823591, "bbox/AP-bicycle": 55.60259086672564, "bbox/AP-boat": 45.743182977295625, "bbox/AP-bottle": 50.22017040664261, "bbox/AP-bus": 77.02453729937041, "bbox/AP-car": 58.43270065158658, "bbox/AP-cat": 49.29044282114153, "bbox/AP-chair": 46.3484178888368, "bbox/AP-cup": 53.925313790400764, "bbox/AP-dining table": 40.141973795164695, "bbox/AP-dog": 56.32446690411162, "bbox/AP-motorcycle": 46.2676739566681, "bbox/AP-person": 51.04507606088663, "bbox/AP50": 81.51711738361045, "bbox/AP75": 58.05245841083272, "bbox/APl": 59.74708471318176, "bbox/APm": 41.10229970000033, "bbox/APs": 14.776018263619003, "iteration": 50000} without featenhaner {"bbox/AP": 52.00824840934456, "bbox/AP-bicycle": 55.64780083617665, "bbox/AP-boat": 44.62704841206179, "bbox/AP-bottle": 48.44416421492211, "bbox/AP-bus": 76.49049907626319, "bbox/AP-car": 58.964925286301096, "bbox/AP-cat": 47.929252741001896, "bbox/AP-chair": 46.04053253887865, "bbox/AP-cup": 53.15243419346582, "bbox/AP-dining table": 39.62968835904925, "bbox/AP-dog": 56.46208801636754, "bbox/AP-motorcycle": 45.92280466010041, "bbox/AP-person": 50.78774257754635, "bbox/AP50": 81.12029339996903, "bbox/AP75": 57.65130580263261, "bbox/APl": 59.240512336751515, "bbox/APm": 40.55028404461523, "bbox/APs": 16.60758248236078, "iteration": 50000}I suspect I didn't train enough and didn't tune the parameters correctly. I would appreciate it if you could give me some suggestions.
bro, I also met that prob. When I trained completely on exdark dataset, only got the 81.4 mAP50. I cant reproduce the performance author provided.
have you solved this prob?Can we exchange from each other through email: pengcaiping0912@gmail.com
Dhs2004
Thank you for your patience and assistance. While comparing FQ R-CNN, I encountered a new issue. I trained both FQ R-CNN and featenhancer+FQ R-CNN on an exdark dataset with different partitioning of training and test sets from https://github.com/cuiziteng/ICCV_MAET, using the same solver configuration. However, I found that the accuracy of the two models is very close. The configuration details are as follows:
I think the featenhaner part is contain in https://github.com/khurramHashmi/FeatEnHancer/blob/33da74366d95ebb9b9a94ca85db99ea94f844308/low-light-object-detection-detectron2/queryrcnn/detector.py#L172
so i test FQ R-CNN in different project cloned from https://github.com/hustvl/Featurized-QueryRCNN. but they shared a similar config file when training(i kept Featurized-QueryRCNN's MODEL part)
result:
I suspect I didn't train enough and didn't tune the parameters correctly. I would appreciate it if you could give me some suggestions.