Operator Parameters
Tip
Pick the default parameters for most cases, and tweak only if your dataset is unusually small, noisy, or large.
Unified Detection (FD + FE)
| Parameter |
Type |
Default |
Description |
target |
- |
dataset/view/selected/current sample |
Scope of operation |
detection_mode |
- |
find_description |
find_description, find_everything, or both |
GPU |
- |
- |
Device picker (cuda:0, cuda:1, etc) |
prompts |
string |
- |
Comma- or newline-separated phrases for VLM |
vlm_mode |
- |
both |
both, florence, or owl |
owl_threshold |
float |
0.30 |
Minimum confidence for OWL-v2 detections |
iou_merge |
float |
0.50 |
IoU threshold to merge Florence-2 + OWL outputs |
nms_iou |
float |
0.60 |
Non-max suppression threshold for merged VLM boxes |
max_area_percentage |
float |
0.20 |
Max object area fraction for SAM (shown when mode is find_everything or both) |
min_area_percentage |
float |
0.003 |
Min object area fraction for SAM (shown when mode is find_everything or both) |
sam_nms_thresh |
float |
0.70 |
NMS threshold for SAM boxes (shown when mode is find_everything or both) |
points_per_batch |
int |
32 |
SAM grid points per batch (shown when mode is find_everything or both) |
overlap_iou_threshold |
float |
0.50 |
Suppresses VLM boxes overlapping existing labeled ones |
batch_size |
int |
4 |
Operator processing batch size |
Outputs
leip_embedding on samples (via SAM encoder)
vlm_detections (intermediate) when VLM runs
unified_detections (final):
- includes new VLM detections (
source='vlm')
- includes new SAM detections (as
label='unknown', source='sam')
- retains existing detections depending on mode
bbox_area computed for every detection and stored as a field expression for patches
- Per-detection
leip_embedding extracted from the image embedding over each bounding box region
Merge semantics
- In
both mode, new VLM boxes that overlap too much (IoU > overlap_iou_threshold) with existing non-unknown detections are skipped to avoid conflicts.
- SAM boxes that overlap an existing detection (per SAM NMS/overlap logic) are not added. Only novel boxes are added as
unknown.
Memory management
- Loads SAM/VLM models only as needed.
- Between stages, aggressively frees GPU memory using an internal cleanup manager (
torch.cuda.empty_cache(), GC, and object deletion). Cleanup summaries are logged.
Find Similar
| Parameter |
Type |
Default |
Description |
target |
- |
dataset/view/selected/current sample |
Scope of operation |
clustering_method |
- |
hybrid |
hybrid = K-Means + majority + optional LDA; kmeans = K-Means only (no LDA) |
target_objs_per_cluster |
int |
15 |
Desired average cluster size |
majority_threshold |
float |
0.55 |
Fraction required to assign cluster majority label to unknowns |
min_classes_for_lda |
int |
3 |
Minimum distinct classes to enable LDA |
min_examples_per_class |
int |
5 |
Minimum examples per class for LDA training |
n_neighbors |
int |
15 |
UMAP neighbors for cluster visualization |
batch_size |
int |
4 |
Processing batch size |
Notes
- Requires
unified_detections and leip_embedding.
- Outputs: updated
label and cluster_id; optional UMAP visualization.
Modify Detections
| Parameter |
Type |
Default |
Description |
action |
- |
change_label |
change_label or remove_detections |
target |
- |
dataset/view/selected/current sample |
Scope of operation |
unknown_only |
bool |
True |
Apply only to unknown detections (safer bulk ops) |
Notes
- Operates on both full samples and patch views.
- Patch removal updates the parent sample's
unified_detections accordingly.