Per-task encoder scripts
Each complex prediction type has its own Python file that reads how predictions are recorded (bbox, polygon, keypoint, etc.) and outputs one image or text file per encoding variant.
| Task | Script | Prediction format | Outputs |
|---|---|---|---|
| depth_estimation | src/encoders/encode_depth.py |
predictions[].url (3 colormap URLs) |
3 images per sample: plasma, turbo, gray |
| object_detection | src/encoders/encode_object_detection.py |
predictions_type: "bbox", predictions[].bbox [x1,y1,x2,y2], label, color_hex |
original, box_only, box_label, text_xyxy.txt, text_xywh.txt |
| instance_segmentation | src/encoders/encode_instance_segmentation.py |
Polygon: predictions[].polygon [[x,y],...]. RLE: predictions[].rle (needs pycocotools to decode) |
original, enc1..enc6 (overlay images), enc7_json.txt |
| semantic_segmentation | src/encoders/encode_semantic_segmentation.py |
predictions[].polygon [[[x,y],...]] or [[x,y],...], label |
same 7 as instance (enc1–7) |
| referring_segmentation | src/encoders/encode_referring_segmentation.py |
predictions[].polygon, label (referring expression) |
original, fill, contour, fill_contour, json.txt |
| keypoint | src/encoders/encode_keypoint.py |
predictions[].keypoint (17×3: x,y,conf per COCO keypoint) |
original, enc_a (same color), enc_b (per-person), enc_c (per-limb), enc_d_json.txt |
| generation_* / lowlevel-* | src/encoders/encode_generation_lowlevel.py |
predictions[].url or predictions[].image (prediction image URL) |
1 image per sample |
Shared: src/common.py provides load_image_id_to_url() (from images/*.json) and fetch_image_cv2(url) for overlay tasks.
Requirement for overlay tasks: images/*.json (or root images.json) must list each image with id and url so that image_id in annotations can be resolved to the original image URL.
Run all: ./run_encoding.sh (uses 24 threads, 2 samples per task).