PseDet:Revisiting the Power of Pseudo Label in Incremental Object Detection

[ICLR 2025] Revisiting the Power of Pseudo Label in Incremental Object Detection

🚀 Overview

• We revisit the strategy of pseudo-labeling on incremental object detection and identify three critical problems that hinder it from achieving competitive performance.
• Based on the above findings, we propose PseDet, a simple yet effective pseudo-labeling framework for incremental object detection, which consists of three key components: spatiotemporal enhancement module, categorical adaptive label selector, and confidence score calibration supervision.
• Extensive experiments conducted on the MS COCO dataset with various incremental settings validate the effectiveness and generalization of our approach. Notably, PseDet outperforms previous methods by 4~17 mAP on different learning settings, achieving new state-of-the-art in incremental object detection.

🛠️ Requirements

Before you begin, ensure you have the following prerequisites installed on your system:

• Python (3.6+)
• PyTorch (1.3+)
• CUDA (9.2+)
• MMDetection.

💻 QuickStart

Two Stage Scenarios

1.Train the initial model using the provided script:

./tools/dist_train.sh ./coco_cfg/train_stage_init/gfl_r50_1x_s0_40cls.py $nums_gpu

2.Infer Pseudo-Labels

Generate pseudo-labels for the first stage:

./tools/dist_test.sh ./coco_cfg/generate_pseudo/gfl_infer_classes_40.py $checkpoint_path $nums_gpu --tta

3.Second Stage (40+40 cls)

Train the model on the second stage with pseudo-labels:

./tools/dist_train.sh ./coco_cfg/train_stage_40_40/gfl_r50_1x_44_s1_80cls.py $nums_gpu

Multi Stage Scenarios

1.Train the initial model (40 cls) and Infer Pseudo-Labels

./tools/dist_train.sh ./coco_cfg/train_stage_init/gfl_r50_1x_s0_40cls.py $nums_gpu
./tools/dist_test.sh ./coco_cfg/generate_pseudo/gfl_infer_classes_40.py $checkpoint_path $nums_gpu --tta

2.Train Second Stage (40+20 cls) and Infer Pseudo-Labels

./tools/dist_test.sh ./coco_cfg/train_stage_40_20_20/gfl_r50_1x_422_s1_60cls.py $nums_gpu
./tools/dist_test.sh ./coco_cfg/generate_pseudo/gfl_infer_classes_40.py $checkpoint_path $nums_gpu --tta

3.Train Third Stage (40+20+20 cls)

./tools/dist_test.sh ./coco_cfg/train_stage_40_20_20/gfl_r50_1x_422_s2_80cls.py $nums_gpu

⚙️ Key Modules

1. Dataset Loader

Multi-stage COCO dataloader, which is capable of loading learning data from different stages during various phases of training.

dataset=dict(
    type='CocoDatasetCLMultiStage,
    data_root=data_root,
    ann_file='annotations/instances_train2017.json',
    pseudo_file=pseudo_file, ## load pseudo labels, None for initial stage
    temporal=temporal, ## perform temporal enhancement
    data_prefix=dict(img='train2017/'),
    filter_cfg=dict(filter_empty_gt=True, min_size=32,
            classes_groups=classes_groups, ## [40,40] for two stage, [40,20,20] for multi stage
            class_spilt=True, ## True for continual learning
            stage_index = stage_index),
    pipeline=train_pipeline,
    backend_args=backend_args)

2.Evaluator

An evaluator designed for continual object detection, suitable for separate evaluation across different learning steps.

dict(
  type='CocoMetricCLMultiStage',
  class_group=classes_groups, ## [40,40] for two stage, [40,20,20] for multi stage
  ...
)

3. Weight Initializer

In 40-40 setting, the number of classes for the head in the first stage is 40, and in the second stage, it's 80.

A special hook is needed to load the weights: for the first 40 classes, we load the weights, while for the parameters of latter 40 classes, we initialize them.

custom_hooks = [
    dict(type='CLLoadCheckpoint')
]

4. Detector

Detector for confidence score calibration supervision.

bbox_head=dict(
  type='GFLHeadPseudoDynamic',
  num_classes=num_classes, ## Head CLASSES
  score_type = 'multiply', ## default: 'multiply'
  proj_score = True ## True for confidence score calibration
  ...
)

5. Spatio-Temporal Enhancement Module

Ehance quality of pseudo labels by spatio sampling and temporal combination during inference.

dict(
type='DetTTAModelScalesKmeans',
tta_cfg=dict(nms=dict(type='nms',iou_threshold=0.4),
              max_per_img=100,
              kmeans=True,
              buffer_size=(num_classes,buffer_size)))

🙏 Acknowledgment

This work is implemented based on mmdet. We greatly appreciate their valuable contributions to the community.

📖 Citation

@inproceedings{wangpsedet,
  title={PseDet: Revisiting the Power of Pseudo Label in Incremental Object Detection},
  author={Wang, Qiuchen and Chen, Zehui and Yang, Chenhongyi and Liu, Jiaming and Li, Zhenyu and Zhao, Feng},
  booktitle={The Thirteenth International Conference on Learning Representations}
}