EgoObjects final detection system
This repository contains the inference weights for the final AIAA 4220 Project 2 ensemble. Seven detectors jointly form one system: DDQ-DETR, YOLOv8x, Co-DETR, D-FINE-X, DINO, Cascade R-CNN and Co-DINO. A small fitted quality head adjusts native DDQ scores. Co-DETR's query, RoI and ATSS heads, and its flipped query view, share one checkpoint.
The reported train-only system achieved Kaggle Public 0.90872 and Private 0.90631. Development-validation mAP is 0.9030299197648789 and AP95 is 0.5353143927561753. Validation was repeatedly used for checkpoint and configuration selection. These metrics belong to the complete ensemble, not any individual checkpoint.
Files
Inference_Weights_01.zip: Co-DETR.Inference_Weights_02.zip: DDQ and DINO.Inference_Weights_03.zip: Co-DINO, Cascade, D-FINE and YOLO.trainonly_quality_head.pth: fitted DDQ quality head; also included in the course code package.WEIGHTS_MANIFEST.json: checkpoint sizes, SHA-256 checksums and source-checkpoint identities.ARCHIVES.json: archive sizes, checksums and contents.
Extract all three archives into one weights directory. They are ordinary independent ZIP files, not binary archive fragments. The duplicate WEIGHTS_MANIFEST.json is identical in all three archives.
Inference
Use the submitted source package's tools/download_weights.py and tools/infer_images.py. The source package supplies resolved detector configurations, model-specific runners, the pinned Co-DETR and D-FINE implementation trees, quality ranking, horizontal-flip merging and the final ensemble postprocessor. The seven models require separate compatible framework environments; these files are not a single Transformers model and cannot be loaded through AutoModel.
The inference runner accepts either the organizer's image index or a directory of new images, creates an annotation-free index, runs every detector and exports COCO-format detections and a prediction CSV. The ten classes are pillow, box, book, bottle, chair, mug, door, shelf, plate and sofa. Images and dataset annotations are not included here.
Checkpoint preparation and validation
Inference tensors were extracted without changing their values or dtypes. Optimizer state, training metadata and unused Co-DETR EMA backup tensors were removed. Original and extracted active tensors were compared exactly. Normal Co-DETR state-dict keys preserve the checkpoint's inference weights; ema_* backup keys are excluded. YOLO retains the selected EMA model as a state dict plus its architecture, and requires the supplied loader.
The complete pipeline for fresh image inference has not been independently revalidated end to end on a GPU. Exact checkpoint extraction and cache-to-output equality do not by themselves establish full-system GPU reproducibility or accuracy on new images. No retraining, quantization or model removal was performed for this release.
Attribution and scope
Framework and architecture sources include MMDetection, Ultralytics, Co-DETR and DEIM. Applicable upstream code, pretrained-weight and dataset terms remain relevant; this repository does not replace them with a blanket license. Source revisions and initializer lineage are documented in the submitted report and model manifest.
All project fitting for the reported system uses the training split. No test annotations were used. Co-DETR inherits Objects365/EVA-02/COCO pretraining; its eligibility under the course rules remains unresolved and is disclosed in the report.
Training resources for course reproduction
Recovered DDQ/DINO full continuation states, Co-DINO/Cascade/D-FINE initializers, and D-FINE training split inputs are available in the pinned training resource release. Check its MANIFEST.json for sizes, original identities and sanitized-file SHA-256 checksums. No login is required.
Use tools/download_training_resources.py and TRAINING_RESOURCES.md in the updated course code supplement. Checkpoint metadata has been sanitized while all tensors were verified unchanged. These training resources are separate from the compact inference archives above. Organizer images remain external. Availability of these states does not establish a full training rerun or exact reproduction of the reported metrics.