MMDetection

MMDetection uses a registry pattern to enable dynamic composition of detection models from interchangeable components (backbone, neck, head, loss). Users configure detectors declaratively via Python config files that instantiate registered modules, allowing researchers to mix-and-match architectures without modifying core framework code. The registry system resolves string identifiers to concrete implementations at runtime, supporting inheritance and override patterns for customization.

MMDetection provides a curated collection of 300+ pre-trained detection models spanning single-stage (YOLO, SSD, RetinaNet), two-stage (Faster R-CNN, Cascade R-CNN), and transformer-based (DINO, Grounding DINO) architectures. Models are trained on standard benchmarks (COCO, LVIS, Objects365) with published metrics and are stored in a unified checkpoint format that includes model weights, config, and metadata. The framework provides utilities to load, validate, and fine-tune these checkpoints with minimal code.

MMDetection provides a high-level inference API (inference_detector function) that loads a model from checkpoint, runs inference on images or batches, and returns predictions in a standardized format. The framework includes visualization utilities that overlay predicted boxes, masks, and class labels on images with configurable colors and transparency. Inference supports both single images and batches with automatic batching and padding.

MMDetection implements test-time augmentation where multiple augmented versions of an image (flips, rotations, scales) are processed through the detector, and predictions are aggregated via NMS or voting. TTA is configured declaratively in the config file and applied during inference without modifying the model. The framework handles coordinate transformation to map predictions from augmented space back to original image space.

MMDetection provides training pipelines for semi-supervised detection (using unlabeled data with pseudo-labels) and weakly-supervised detection (using image-level labels instead of box annotations). The framework includes utilities for pseudo-label generation, confidence filtering, and auxiliary losses that leverage unlabeled data. Semi-supervised training alternates between supervised and unsupervised phases with configurable pseudo-label thresholds.

MMDetection provides analysis tools for understanding detector behavior: feature map visualization (showing what features the model learns), attention map visualization (for transformer-based detectors), prediction analysis (false positives, false negatives, localization errors), and dataset statistics. These tools help practitioners debug poor performance by identifying failure modes (e.g., small object detection failures, class confusion).

MMDetection implements a structured data processing pipeline where image augmentation, normalization, and annotation transforms are defined declaratively in config files as a sequence of composable operations. Each transform (Resize, RandomFlip, Normalize, etc.) is a registered class that processes both images and bounding box/segmentation annotations consistently. The pipeline is executed during dataset iteration, with transforms applied in order and supporting both training (with augmentation) and inference (without) modes.

MMDetection provides dataset adapters that normalize diverse annotation formats (COCO JSON, Pascal VOC XML, LVIS, Objects365, custom formats) into a unified internal representation. The framework includes a dataset registry where users register custom dataset classes that implement a standard interface (load annotations, get image/label pairs). During training, the framework can mix multiple datasets via weighted sampling or sequential batching, with automatic format conversion and validation.

MMDetection implements single-stage detectors as end-to-end models that predict bounding boxes and class scores directly from feature maps without region proposal generation. The framework provides modular implementations of YOLO, SSD, RetinaNet, and ATSS architectures with configurable backbones (ResNet, ResNeXt, Swin), necks (FPN, PAFPN), and heads. Single-stage detectors use dense prediction heads that output predictions at multiple scales, with focal loss or other loss functions to handle class imbalance.

MMDetection implements two-stage detectors that first generate region proposals via RPN (Region Proposal Network), then refine predictions in a second stage using ROI pooling/alignment. The framework provides modular implementations of Faster R-CNN, Cascade R-CNN (with iterative refinement), and Mask R-CNN (with instance segmentation). Two-stage detectors use separate classification and bounding box regression heads per proposal, enabling higher accuracy especially on small objects.

MMDetection implements transformer-based detectors that replace hand-crafted features and anchors with learned attention mechanisms. The framework provides implementations of DINO (Detection with Implicit Oriented windows), Grounding DINO (open-vocabulary detection with text grounding), and DETR variants. These detectors use transformer encoders/decoders to process multi-scale features, with learnable query embeddings that attend to image features to predict boxes and classes.

MMDetection provides a modular loss function system where classification, localization, and auxiliary losses are registered and composed via config. The framework includes focal loss (for class imbalance), IoU-based losses (GIoU, DIoU, CIoU), L1/smooth-L1 losses, and task-specific losses (mask loss, keypoint loss). Loss functions are weighted and combined during training, with support for dynamic loss weighting and auxiliary losses for semi-supervised learning.

MMDetection provides distributed training support via PyTorch's DistributedDataParallel (DDP) with automatic gradient synchronization across GPUs/nodes. The framework handles batch size scaling, learning rate adjustment, and gradient accumulation for distributed settings. Training is coordinated via a config-driven launcher that manages process spawning, rank assignment, and checkpoint synchronization across workers.

MMDetection provides evaluation utilities that compute standard detection metrics (mean Average Precision at IoU thresholds, Average Recall) following COCO and LVIS evaluation protocols. The framework includes metric implementations for bounding box detection, instance segmentation, panoptic segmentation, and rotated object detection. Evaluation is performed on validation/test sets with configurable IoU thresholds and class-specific breakdowns.