
Finds object-detection images where a model's predictions miss ground truth per class
- Open source
- 4K GitHub stars
- Computer vision

Finds object-detection images where a model's predictions miss ground truth per class

Finds the weakest VCN classify predictions and ranks them for augmentation via TAO

Finds false-positive and false-negative yes/no VLM predictions for root-cause analysis

Validates TAO job artifacts: spec bundles, job records, results layout and best_rec

Moves TAO job data to and from compute containers, choosing the storage tier first

Prepares and validates a dataset and recipe for fine-tuning AnomalyGenNext LoRA weights

Guides fine-tuning of NVIDIA Cosmos3 reasoning models with Cosmos Framework or Cosmos-RL

Computes image embeddings from a parquet of file paths using CLIP, SigLIP or TAO

Generates synthetic object-detection defect images with COCO labels via AnomalyGenNext

Turns raw videos into chain-of-thought QA training data using VLM and LLM teacher models

Mines look-alike source images for VCN augmentation via DEFT embedding

Mines nearest-neighbor source samples for target embeddings using TAO Data Services

Mines unique source images closest to target samples for object detection datasets

Prepares false-negative masks and clean-image inputs for AnomalyGenNext generation runs

Routes weak VCN samples into per-module parquet subsets for mining and AnomalyGen

Runs container-backed hyperparameter search for NVIDIA TAO networks via AutoMLRunner

Runs an iterative NVIDIA TAO AOI loop that improves PCB defect-inspection models

Runs an iterative DEFT loop that improves Cosmos3 bare OK/NG AOI inspection accuracy

Runs an iterative mining-based DEFT loop to improve Cosmos Reason traffic-video

Runs NVIDIA TAO Grounding DINO detection through an iterative data-mining and retraining

Runs an iterative RT-DETR retraining loop for binary industrial defect detection

Runs an iterative retrain-and-evaluate loop to improve NVIDIA TAO CLIP image-text

Lets an agent submit, monitor and cancel TAO GPU training jobs on Kubernetes clusters

Runs Python training and eval scripts inside a local virtualenv, without Docker

Prepares a session to run NVIDIA TAO skills: host checks, credential checks, and workflow

Checks and installs the NVIDIA driver, CUDA and container toolkit for TAO GPU hosts

Runs staged TAO workflows as small fresh agent sessions to cut token use

Trains, evaluates and exports Siamese models that flag defects in AOI image pairs

Trains, evaluates and exports TAO Sparse4D multi-camera 3D detection and tracking models

Trains, evaluates and runs Visual ChangeNet models for PCB and AOI defect inspection