{"passport":{"unfragile":{"@version":"1.0","version":"2026-05","artifact":{"id":"pypi_pypi-mmdet","slug":"pypi-mmdet","name":"mmdet","type":"benchmark","url":"https://github.com/open-mmlab/mmdetection","page_url":"https://unfragile.ai/pypi-mmdet","categories":["model-training","testing-quality"],"tags":["computer","vision","object","detection"],"pricing":{"model":"open_source","free":true,"starting_price":null},"status":"active","verified":false},"capabilities":[{"id":"pypi_pypi-mmdet__cap_0","uri":"capability://tool.use.integration.modular.detector.architecture.composition.via.registry.system","name":"modular detector architecture composition via registry system","description":"MMDetection decomposes object detection into pluggable components (backbone, neck, head, loss) registered in a centralized registry pattern, enabling users to construct custom detectors by combining pre-built modules without modifying core framework code. The registry system maps string identifiers to component classes, allowing configuration-driven model instantiation where backbone (ResNet, Swin), neck (FPN, PAFPN), and head (detection, mask, ROI) modules are swapped declaratively.","intents":["I want to experiment with different backbone architectures without rewriting detector code","I need to combine a custom feature pyramid with an existing detection head","I want to add a new loss function and have it automatically available across all detectors"],"best_for":["computer vision researchers prototyping detection architectures","teams building production detection systems with evolving requirements","practitioners extending MMDetection with proprietary components"],"limitations":["Registry-based composition adds indirection layer requiring understanding of component interfaces and contracts","Tight coupling between component input/output shapes can cause silent failures if incompatible modules are combined","Limited compile-time validation of component compatibility — errors surface only at runtime during forward pass"],"requires":["PyTorch 1.9+","Python 3.7+","mmcv library (OpenMMLab common vision library)","Understanding of detector component interfaces (backbone output channels, neck input/output specs)"],"input_types":["configuration files (Python or YAML)","pre-trained model weights"],"output_types":["instantiated PyTorch nn.Module detector","model graph with registered components"],"categories":["tool-use-integration","architecture-composition"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"pypi_pypi-mmdet__cap_1","uri":"capability://automation.workflow.configuration.driven.training.pipeline.with.distributed.support","name":"configuration-driven training pipeline with distributed support","description":"MMDetection abstracts the entire training workflow (data loading, augmentation, optimization, checkpointing) into declarative Python configuration files that specify dataset paths, model architecture, learning rates, schedules, and distributed training parameters. The framework parses these configs and orchestrates multi-GPU/multi-node training via PyTorch DistributedDataParallel, handling gradient synchronization, checkpoint saving, and metric logging automatically without requiring manual distributed training code.","intents":["I want to train a detector on multiple GPUs without writing distributed training boilerplate","I need to reproduce a published detection model with exact hyperparameters from a config file","I want to sweep learning rates and batch sizes by modifying config files, not Python code"],"best_for":["ML engineers training production detection models at scale","researchers reproducing published detection benchmarks","teams managing multiple concurrent training experiments with different hyperparameters"],"limitations":["Configuration files can become deeply nested and hard to debug when combining many modules","Limited support for dynamic/conditional logic in configs — complex training schedules require custom hooks","Distributed training assumes homogeneous hardware; mixed GPU types or heterogeneous clusters require manual tuning"],"requires":["PyTorch 1.9+","CUDA 10.2+ for GPU training","mmcv library","NCCL library for multi-GPU communication","Properly formatted dataset (COCO, Pascal VOC, or custom with annotations)"],"input_types":["Python configuration files (.py)","image datasets with annotation files","pre-trained backbone weights"],"output_types":["trained model checkpoints (.pth)","training logs and metrics (JSON, TensorBoard)","evaluation results on validation set"],"categories":["automation-workflow","planning-reasoning"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"pypi_pypi-mmdet__cap_10","uri":"capability://data.processing.analysis.semi.supervised.object.detection.with.pseudo.labeling.and.consistency.regularization","name":"semi-supervised object detection with pseudo-labeling and consistency regularization","description":"MMDetection supports semi-supervised detection where unlabeled data is leveraged via pseudo-labeling (generating predictions on unlabeled data and using high-confidence predictions as training targets) and consistency regularization (enforcing consistent predictions under different augmentations). The framework implements teacher-student models where a teacher network generates pseudo-labels for unlabeled data, and a student network is trained on both labeled and pseudo-labeled data with consistency losses.","intents":["I want to leverage unlabeled data to improve detection accuracy when labeled data is limited","I need to implement pseudo-labeling where high-confidence predictions on unlabeled data become training targets","I want to use consistency regularization to enforce stable predictions under augmentation"],"best_for":["teams with large unlabeled datasets and limited labeled data","practitioners improving detection accuracy in low-data regimes","researchers exploring semi-supervised learning for detection"],"limitations":["Pseudo-labeling quality depends on teacher model confidence — incorrect pseudo-labels can degrade performance","Teacher-student training requires careful tuning of pseudo-label confidence thresholds and consistency loss weights","Semi-supervised training is more complex than supervised training, requiring monitoring of pseudo-label quality","Consistency regularization adds computational overhead due to multiple forward passes per sample"],"requires":["PyTorch 1.9+","mmcv library","Large unlabeled dataset (typically 10-100x larger than labeled data)","Labeled dataset for initial teacher training","GPU with 32GB+ memory for teacher-student training"],"input_types":["labeled images with bounding box annotations","unlabeled images (no annotations)"],"output_types":["trained detector with improved accuracy from semi-supervised learning","pseudo-label quality metrics (confidence distribution, recall on unlabeled data)"],"categories":["data-processing-analysis","planning-reasoning"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"pypi_pypi-mmdet__cap_11","uri":"capability://data.processing.analysis.model.analysis.and.visualization.tools.for.debugging.and.interpretation","name":"model analysis and visualization tools for debugging and interpretation","description":"MMDetection provides analysis tools for visualizing model predictions, attention maps, and feature activations to aid debugging and interpretation. The framework includes visualization utilities for drawing bounding boxes, segmentation masks, and attention heatmaps on images, as well as analysis tools for computing prediction confidence distributions, false positive/negative analysis, and per-class performance breakdown. These tools help practitioners understand model behavior and identify failure modes.","intents":["I want to visualize model predictions on test images to identify failure modes","I need to analyze per-class detection performance to find weak classes","I want to visualize attention maps from transformer-based detectors to understand what regions the model focuses on"],"best_for":["practitioners debugging detection models and identifying failure modes","researchers analyzing model behavior and attention patterns","teams conducting error analysis to prioritize improvements"],"limitations":["Visualization tools are primarily for offline analysis — not suitable for real-time monitoring","Attention visualization from transformer models can be ambiguous due to multi-head attention","False positive/negative analysis requires ground truth annotations, limiting applicability to test sets","Visualization tools don't provide actionable insights — require manual interpretation"],"requires":["PyTorch 1.9+","mmcv library","matplotlib or similar visualization library","Trained detector checkpoint","Test images with optional ground truth annotations"],"input_types":["images","model predictions (bounding boxes, confidence scores, masks)","ground truth annotations (optional)"],"output_types":["visualized images with predictions overlaid","attention heatmaps","performance analysis plots (confidence distribution, per-class AP)","false positive/negative examples"],"categories":["data-processing-analysis","safety-moderation"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"pypi_pypi-mmdet__cap_2","uri":"capability://data.processing.analysis.multi.stage.data.augmentation.pipeline.with.geometric.and.photometric.transforms","name":"multi-stage data augmentation pipeline with geometric and photometric transforms","description":"MMDetection provides a composable data augmentation pipeline that applies geometric transforms (resize, crop, rotate, flip) and photometric transforms (color jitter, normalization) in sequence, with bounding box and segmentation mask updates automatically propagated through each transform. The pipeline is defined declaratively in config files and supports both online augmentation (applied during training) and test-time augmentation (TTA) where multiple augmented versions of test images are inferred and results are aggregated.","intents":["I want to apply consistent augmentation to images and their bounding box annotations without manual coordinate updates","I need to implement test-time augmentation by inferring on multiple augmented versions of test images","I want to compose custom augmentation sequences (e.g., mosaic, mixup) with standard transforms"],"best_for":["computer vision practitioners training robust detection models with limited data","teams implementing advanced augmentation strategies (mosaic, mixup, CutMix)","researchers evaluating detection model robustness via test-time augmentation"],"limitations":["Augmentation pipeline is applied sequentially, which can be slow for large batches — no GPU-accelerated augmentation by default","Complex augmentations (mosaic, mixup) require careful tuning of parameters to avoid corrupting annotations","Test-time augmentation increases inference latency linearly with number of augmented versions (e.g., 5x slower for 5 TTA variants)"],"requires":["PyTorch 1.9+","mmcv library","Pillow or OpenCV for image I/O","numpy for coordinate transformations","Properly formatted bounding box annotations (COCO format recommended)"],"input_types":["images (JPEG, PNG, etc.)","bounding box annotations (COCO JSON, Pascal VOC XML, or custom)","segmentation masks (PNG, RLE encoded)"],"output_types":["augmented images with transformed annotations","aggregated predictions from multiple TTA variants"],"categories":["data-processing-analysis","image-visual"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"pypi_pypi-mmdet__cap_3","uri":"capability://image.visual.single.stage.detector.implementation.yolo.ssd.retinanet.atss.variants","name":"single-stage detector implementation (yolo, ssd, retinanet, atss variants)","description":"MMDetection provides implementations of single-stage detectors that predict bounding boxes and class scores directly from feature maps without region proposal generation. These detectors use dense prediction heads that output predictions at multiple scales (via FPN), with focal loss to handle class imbalance and IoU-based loss functions for box regression. The architecture supports anchor-based (YOLO, SSD, RetinaNet) and anchor-free (FCOS, ATSS) variants with configurable backbone and neck modules.","intents":["I want to train a fast, single-stage detector for real-time inference on edge devices","I need to implement a custom single-stage detector variant by modifying the detection head","I want to compare anchor-based vs anchor-free detection approaches on my dataset"],"best_for":["practitioners building real-time detection systems with latency constraints","researchers experimenting with single-stage detector architectures","teams deploying detectors on mobile/edge hardware with limited compute"],"limitations":["Single-stage detectors typically have lower accuracy than two-stage detectors on small objects due to limited receptive field","Anchor-based variants require careful tuning of anchor scales and aspect ratios per dataset","Focal loss and other class imbalance techniques add training complexity and hyperparameter tuning burden"],"requires":["PyTorch 1.9+","mmcv library","COCO or custom detection dataset with bounding box annotations","GPU with sufficient memory for batch training (8GB+ recommended)"],"input_types":["images (any resolution, auto-resized during training)","bounding box annotations with class labels"],"output_types":["predicted bounding boxes with confidence scores","class predictions per box","inference speed (FPS) metrics"],"categories":["image-visual","planning-reasoning"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"pypi_pypi-mmdet__cap_4","uri":"capability://image.visual.two.stage.detector.implementation.faster.r.cnn.cascade.r.cnn.mask.r.cnn.variants","name":"two-stage detector implementation (faster r-cnn, cascade r-cnn, mask r-cnn variants)","description":"MMDetection implements two-stage detectors that first generate region proposals (via RPN) and then refine them with classification and bounding box regression heads. The framework supports cascade refinement (Cascade R-CNN) where proposals are progressively refined through multiple stages with increasing IoU thresholds, and instance segmentation (Mask R-CNN) where a mask head predicts per-pixel segmentation masks for each detected instance. ROI pooling/alignment extracts fixed-size features from proposals for downstream processing.","intents":["I want to train a high-accuracy detector that handles small objects and occlusion better than single-stage methods","I need to perform instance segmentation alongside object detection","I want to implement cascade refinement to progressively improve proposal quality"],"best_for":["practitioners prioritizing detection accuracy over inference speed","teams building instance segmentation systems","researchers implementing multi-stage detection refinement strategies"],"limitations":["Two-stage detectors are 2-5x slower than single-stage methods due to RPN + refinement stages","RPN requires careful tuning of anchor scales, aspect ratios, and NMS thresholds","Cascade R-CNN training is more complex due to multiple refinement stages with different IoU thresholds","Mask head adds significant memory overhead during training (requires storing masks for all proposals)"],"requires":["PyTorch 1.9+","mmcv library","COCO dataset (or custom with instance segmentation masks for Mask R-CNN)","GPU with 16GB+ memory for training with large batch sizes"],"input_types":["images (any resolution)","bounding box annotations with class labels","instance segmentation masks (PNG or RLE encoded) for Mask R-CNN"],"output_types":["predicted bounding boxes with confidence scores","class predictions per box","instance segmentation masks (for Mask R-CNN)","proposal quality metrics (recall@k)"],"categories":["image-visual","planning-reasoning"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"pypi_pypi-mmdet__cap_5","uri":"capability://image.visual.transformer.based.detector.implementation.detr.deformable.detr.dino.variants","name":"transformer-based detector implementation (detr, deformable detr, dino variants)","description":"MMDetection provides implementations of transformer-based detectors (DETR, Deformable DETR, DINO) that replace hand-crafted detection heads with learned transformer encoders/decoders. These detectors treat object detection as a set prediction problem where a fixed number of learnable query embeddings are refined through transformer layers to predict bounding boxes and class scores. Deformable attention mechanisms enable efficient processing of high-resolution feature maps by attending only to relevant spatial regions.","intents":["I want to train a transformer-based detector that doesn't require hand-crafted anchors or NMS","I need to implement deformable attention for efficient multi-scale feature processing","I want to leverage pre-trained vision transformers (ViT, Swin) as backbones for detection"],"best_for":["researchers exploring transformer architectures for detection","teams leveraging pre-trained vision transformers for downstream detection tasks","practitioners building end-to-end differentiable detection systems without NMS"],"limitations":["Transformer-based detectors require significantly more training iterations (500+ epochs) to converge compared to CNN-based detectors (12 epochs)","Deformable attention adds complexity and requires careful tuning of attention offset learning","Set prediction loss (Hungarian matching) is computationally expensive for large numbers of queries","Inference speed is comparable to two-stage detectors despite simpler architecture due to transformer overhead"],"requires":["PyTorch 1.9+","mmcv library","timm library for pre-trained vision transformer backbones","GPU with 32GB+ memory for training large transformer models","COCO or large-scale detection dataset (transformers require more data than CNNs)"],"input_types":["images (any resolution, typically 800-1333 pixels)","bounding box annotations with class labels","pre-trained vision transformer weights (optional)"],"output_types":["predicted bounding boxes with confidence scores","class predictions per box","attention visualizations (optional)"],"categories":["image-visual","planning-reasoning"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"pypi_pypi-mmdet__cap_6","uri":"capability://image.visual.multi.task.learning.with.panoptic.and.instance.segmentation.heads","name":"multi-task learning with panoptic and instance segmentation heads","description":"MMDetection supports multi-task learning where detection, instance segmentation, and panoptic segmentation are trained jointly with shared backbones and necks. The framework provides separate heads for each task (detection head, mask head, semantic segmentation head) that operate on shared feature maps, with task-specific losses combined via weighted summation. Panoptic segmentation unifies instance and semantic segmentation by assigning each pixel to either an instance or semantic class.","intents":["I want to train a single model that performs detection, instance segmentation, and semantic segmentation simultaneously","I need to implement panoptic segmentation by combining instance and semantic predictions","I want to share backbone features across multiple detection-related tasks to reduce model size"],"best_for":["teams building comprehensive scene understanding systems","practitioners with datasets containing both instance and semantic annotations","researchers exploring multi-task learning for detection and segmentation"],"limitations":["Multi-task learning requires careful balancing of task-specific losses to prevent one task from dominating training","Panoptic segmentation requires both instance and semantic annotations, limiting dataset availability","Increased model complexity and training time due to multiple heads and loss functions","Panoptic quality metrics (PQ) are harder to interpret than individual instance/semantic metrics"],"requires":["PyTorch 1.9+","mmcv library","Datasets with instance segmentation masks AND semantic segmentation labels (e.g., COCO Panoptic)","GPU with 32GB+ memory for training multi-task models"],"input_types":["images","bounding box annotations with class labels","instance segmentation masks (PNG or RLE encoded)","semantic segmentation labels (PNG)"],"output_types":["predicted bounding boxes with confidence scores","instance segmentation masks","semantic segmentation predictions","panoptic segmentation (unified instance + semantic predictions)"],"categories":["image-visual","data-processing-analysis"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"pypi_pypi-mmdet__cap_7","uri":"capability://data.processing.analysis.model.evaluation.with.coco.lvis.and.custom.metrics","name":"model evaluation with coco, lvis, and custom metrics","description":"MMDetection provides comprehensive evaluation metrics for object detection including COCO Average Precision (AP), LVIS metrics (with long-tail class weighting), and custom metrics. The evaluation pipeline computes metrics at multiple IoU thresholds (0.5:0.95), object sizes (small, medium, large), and supports both standard evaluation and class-wise breakdown. Metrics are computed on validation sets during training and on test sets for final model evaluation.","intents":["I want to evaluate my detector using standard COCO metrics (AP, AP50, AP75, etc.)","I need to evaluate on long-tail datasets (LVIS) where class distribution is imbalanced","I want to analyze per-class detection performance to identify weak classes"],"best_for":["practitioners benchmarking detection models against published baselines","teams evaluating detectors on long-tail datasets with imbalanced class distributions","researchers analyzing detection performance across object sizes and classes"],"limitations":["COCO metrics are computationally expensive for large test sets (requires computing pairwise IoU for all predictions)","LVIS evaluation requires specific annotation format and class grouping (base/novel/rare), limiting applicability to custom datasets","Metrics don't capture failure modes (e.g., false positives on specific object types) — requires manual analysis","AP metric is sensitive to NMS threshold and confidence score calibration"],"requires":["PyTorch 1.9+","mmcv library","pycocotools library for COCO metric computation","Predictions in COCO JSON format (list of {image_id, category_id, bbox, score})","Ground truth annotations in COCO JSON format"],"input_types":["predicted bounding boxes with confidence scores (COCO JSON format)","ground truth annotations (COCO JSON or LVIS JSON format)"],"output_types":["AP (Average Precision) at IoU=0.5:0.95","AP50, AP75 (AP at specific IoU thresholds)","APsmall, APmedium, APlarge (AP by object size)","per-class AP breakdown","recall at different IoU thresholds"],"categories":["data-processing-analysis","safety-moderation"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"pypi_pypi-mmdet__cap_8","uri":"capability://automation.workflow.model.inference.and.deployment.with.batch.processing.and.tta","name":"model inference and deployment with batch processing and tta","description":"MMDetection provides inference APIs that support single-image and batch inference with automatic preprocessing (resizing, normalization) and postprocessing (NMS, score thresholding). The framework supports test-time augmentation (TTA) where multiple augmented versions of input images are inferred and predictions are aggregated via NMS or weighted averaging. Inference can be executed on CPU or GPU with configurable batch sizes for throughput optimization.","intents":["I want to run inference on a trained detector with automatic preprocessing and NMS","I need to perform test-time augmentation to improve detection accuracy at the cost of inference latency","I want to batch multiple images for efficient GPU inference"],"best_for":["practitioners deploying trained detectors in production systems","teams optimizing inference throughput with batch processing","researchers evaluating detector robustness via test-time augmentation"],"limitations":["Batch inference requires images to be resized to the same dimensions, which may distort aspect ratios","Test-time augmentation increases latency linearly with number of augmented versions (e.g., 5x slower for 5 TTA variants)","NMS is a sequential operation that doesn't parallelize well, becoming a bottleneck for large numbers of detections","Inference API doesn't support streaming/online inference — requires buffering entire batches"],"requires":["PyTorch 1.9+","mmcv library","Trained detector checkpoint (.pth file)","Input images (JPEG, PNG, or numpy arrays)","GPU optional but recommended for throughput"],"input_types":["single image or batch of images (numpy arrays, file paths, or tensors)","image metadata (original size, aspect ratio)"],"output_types":["predicted bounding boxes with confidence scores","class predictions per box","inference latency (ms per image)"],"categories":["automation-workflow","image-visual"],"confidence":0.5,"matches":0,"success_rate":0},{"id":"pypi_pypi-mmdet__cap_9","uri":"capability://image.visual.grounded.object.detection.with.text.image.alignment.glip.grounding.dino","name":"grounded object detection with text-image alignment (glip, grounding dino)","description":"MMDetection implements grounded object detection models (GLIP, Grounding DINO) that align image regions with natural language descriptions, enabling detection of arbitrary object classes without training-time class labels. These models use vision-language pre-training where image patches are aligned with text embeddings, allowing zero-shot detection by matching image features to arbitrary text queries. The framework supports both phrase-level grounding (detecting specific noun phrases) and image-level grounding (detecting all objects matching a description).","intents":["I want to detect arbitrary object classes using natural language descriptions without retraining","I need to perform zero-shot detection on novel classes not seen during training","I want to ground specific text phrases to image regions for visual question answering"],"best_for":["practitioners building flexible detection systems that adapt to new object classes via text","teams performing zero-shot detection on novel classes","researchers exploring vision-language models for detection tasks"],"limitations":["Grounded detection requires pre-trained vision-language models (CLIP, ALIGN) which are computationally expensive","Text-image alignment quality depends on pre-training data; models may struggle with domain-specific or rare object descriptions","Inference latency is higher than standard detectors due to text encoding and cross-modal matching","Performance degrades significantly for fine-grained object categories or complex spatial relationships"],"requires":["PyTorch 1.9+","mmcv library","Pre-trained vision-language model (CLIP, ALIGN, or similar)","Text encoder (BERT or similar) for encoding object descriptions","GPU with 16GB+ memory for inference"],"input_types":["images","text descriptions or queries (natural language phrases or sentences)"],"output_types":["predicted bounding boxes for regions matching text descriptions","confidence scores indicating alignment between image regions and text","text-image similarity scores"],"categories":["image-visual","memory-knowledge"],"confidence":0.5,"matches":0,"success_rate":0}],"trust":{"score":30,"verified":false,"data_access_risk":"low","permissions":["PyTorch 1.9+","Python 3.7+","mmcv library (OpenMMLab common vision library)","Understanding of detector component interfaces (backbone output channels, neck input/output specs)","CUDA 10.2+ for GPU training","mmcv library","NCCL library for multi-GPU communication","Properly formatted dataset (COCO, Pascal VOC, or custom with annotations)","Large unlabeled dataset (typically 10-100x larger than labeled data)","Labeled dataset for initial teacher training"],"failure_modes":["Registry-based composition adds indirection layer requiring understanding of component interfaces and contracts","Tight coupling between component input/output shapes can cause silent failures if incompatible modules are combined","Limited compile-time validation of component compatibility — errors surface only at runtime during forward pass","Configuration files can become deeply nested and hard to debug when combining many modules","Limited support for dynamic/conditional logic in configs — complex training schedules require custom hooks","Distributed training assumes homogeneous hardware; mixed GPU types or heterogeneous clusters require manual tuning","Pseudo-labeling quality depends on teacher model confidence — incorrect pseudo-labels can degrade performance","Teacher-student training requires careful tuning of pseudo-label confidence thresholds and consistency loss weights","Semi-supervised training is more complex than supervised training, requiring monitoring of pseudo-label quality","Consistency regularization adds computational overhead due to multiple forward passes per sample","builder identity is not verified yet","no observed match outcomes yet"],"rank_breakdown":{"adoption":0.05,"quality":0.34,"ecosystem":0.62,"match_graph":0.25,"freshness":0.52,"weights":{"adoption":0.25,"quality":0.35,"ecosystem":0.15,"match_graph":0.2,"freshness":0.05}},"observed_outcomes":{"matches":0,"success_rate":0,"avg_confidence":0,"top_intents":[],"last_matched_at":null},"maintenance":{"status":"active","updated_at":"2026-06-17T09:51:05.295Z","last_scraped_at":"2026-05-03T15:20:19.404Z","last_commit":null},"community":{"stars":null,"forks":null,"weekly_downloads":null,"model_downloads":null,"model_likes":null}},"distribution":{"claim_url":"https://unfragile.ai/submit?claim=pypi-mmdet","compare_url":"https://unfragile.ai/compare?artifact=pypi-mmdet"}},"signature":"jTM8ibM2U8g5f+go1kqNNT3QoM1Qi4RAznnBYFBKUxwHeelzsifcoWs02Ummp8H/UnuHXu6ZqLhlCgHs4Xn+AA==","signedAt":"2026-06-21T01:46:29.894Z","signedBy":"unfragile.ai","version":1},"_links":{"self":"https://unfragile.ai/api/v1/passport/pypi-mmdet","artifact":"https://unfragile.ai/pypi-mmdet","verify":"https://unfragile.ai/api/v1/verify?slug=pypi-mmdet","publicKey":"https://unfragile.ai/api/v1/trust-passport-public-key","spec":"https://unfragile.ai/trust","schema":"https://unfragile.ai/schema.json","docs":"https://unfragile.ai/docs"}}