delete_datasets.md

# Dataset Deletion Manifest — CCTV People Counter
# Generated: 2026-09-06
# Purpose: log every dataset folder being removed from local disk, what was done with it,
#          and whether it is needed for future workflows (training from scratch, fine-tuning,
#          inference / Phase 5).

---

## Tier 1 — KEPT locally + uploaded to Hugging Face Hub
### Rationale: small, critical assets required for inference (Phase 5) and any future
###           fine-tuning or resume. Upload before local deletion of Tier 3.

### 1.1  Fine-tuned model weights (trained detector)

| Path | Approx size | What we did with it | Future need | Status |
|---|---|---|---|---|
| models/finetuned/person_detector_v1/train_extended/weights/best.pt | ~24 MB | Best checkpoint from 47-epoch extended run (mot16+COCO combined). mAP50 ≈ 0.739. Used for person detection in Phase 5 pipeline. | **Required** for Phase 5 inference. Also usable as a starting point for future fine-tuning on real venue data (transfer learning). | KEEP + UPLOAD |
| models/finetuned/person_detector_v1/train_extended/weights/last.pt | ~24 MB | Most recent epoch checkpoint from the same 47-epoch run. | Useful if you want to resume training later (e.g. more epochs on real venue data). Optional if you only need inference. | KEEP (optional UPLOAD) |
| models/finetuned/person_detector_v1/train/weights/best.pt | ~6 MB | Best checkpoint from the earlier Phase 4 run (29 epochs, mAP50 ≈ 0.205 then 0.709 at epoch 1). | Lower-value backup; useful for comparison or as a fallback if the extended run is somehow corrupted. Not required for Phase 5. | KEEP + UPLOAD (optional) |

### 1.2  Pretrained base weights

| Path | Approx size | What we did with it | Future need | Status |
|---|---|---|---|---|
| models/pretrained/yolov8n.pt | ~6 MB | Official YOLOv8n base weights used as the starting point for fine-tuning (Phase 4 / 4b). | Required if you ever fine-tune from scratch again (new dataset needs a base checkpoint). Not needed for Phase 5 inference (which uses your fine-tuned best.pt). | KEEP + UPLOAD |

### 1.3  Config files

| Path | Approx size | What we did with it | Future need | Status |
|---|---|---|---|---|
| config/counting_config.yaml | ~? KB | Line-zone / counting config for Phase 5–6 pipeline. | Required for Phase 5 pipeline runs. | KEEP + UPLOAD |
| config/exclusion_config.yaml | ~? KB | Zone exclusion config for Phase 7. | Required for future zone-exclusion work. | KEEP + UPLOAD |

### 1.4  Dataset YAML definitions (small metadata files describing dataset layout)

| Path | Approx size | What we did with it | Future need | Status |
|---|---|---|---|---|
| data/annotated/dataset.yaml | 378 B | Original MOT16 dataset YAML pointing to images/train + images/val. | Defines how the original MOT16 split was laid out. Useful reference when re-ingesting or verifying dataset structure later. | KEEP + UPLOAD |
| data/annotated/dataset_coco.yaml | 624 B | COCO person-only dataset YAML. | Reference for COCO-derived labels. Useful if you later add more COCO data or re-create the combined dataset. | KEEP + UPLOAD |
| data/annotated/dataset_combined.yaml | 485 B | Combined dataset YAML (MOT16 + COCO). **Note:** the paths inside may reference folders that no longer exist after deletion; keep as a record of the intended layout. | Reference. Not directly usable for training after raw datasets are deleted, but documents the combined schema. | KEEP + UPLOAD |

### 1.5  Scripts (pipeline / conversion / training)

| Path | Approx size | What we did with it | Future need | Status |
|---|---|---|---|---|
| scripts/fine_tune.py | ~? KB | Fine-tuning orchestrator. Handles resume (last.pt detection), epoch checkpoint naming (train/ vs train_extended/), LR, patience, etc. | Required if you ever resume/repeat fine-tuning. | KEEP + UPLOAD |
| scripts/create_combined_dataset.py | ~? KB | Merges MOT16 train + COCO person + COCO background into combined train set (88,165 images). | Required if you rebuild the combined dataset from raw sources in the future. | KEEP + UPLOAD |
| scripts/coco_to_yolo.py | ~? KB | Converts COCO annotations (person class only) to YOLO txt format; also samples background-only images. | Required if you process new COCO-derived data. | KEEP + UPLOAD |
| scripts/regen_mot16_val_labels.py | ~? KB | Regenerates MOT16 val labels from GT files (used to fix broken symlinks). | Useful if MOT16 val labels ever go missing. | KEEP + UPLOAD |
| scripts/pets_to_yolo.py | ~? KB | Converts PETS 2009 annotations to YOLO format. | Useful if you ever re-add PETS data. | KEEP + UPLOAD |

---

## Tier 2 — KEPT locally + uploaded to Hugging Face Hub
### Rationale: the actual training corpus used (or intended) for fine-tuning. Back up in
###           full before deleting raw sources, so future re-training is possible without
###           re-downloading 20+ GB of public data.

### 2.1  Annotated training corpus (processed images + labels)

| Path | Approx size | File count | What we did with it | Future need | Status |
|---|---|---|---|---|---|
| data/annotated/images/ (all subdirs) | ~11 GB | ~68,903 files (incl. duplicates) | Processed/symlinked copies of MOT16 train+val, COCO person, COCO background, PETS, and empty venue-frames. Fed to the YOLOv8 trainer as the image source in Phase 4 / 4b. | **Needed if you ever retrain or fine-tune again from scratch on the same corpus** (or a superset). NOT needed for Phase 5 inference (only best.pt is needed). | KEEP locally + UPLOAD to HF (full tree, ~11 GB) |
| data/annotated/labels/ (all subdirs) | ~0.05 GB | ~178,194 files | YOLO-format txt labels for the annotated images. | Paired with images above; same rationale. Critical for any future training on this corpus. | KEEP + UPLOAD |

### 2.2  Subfolder breakdown (for reference — upload as one tree, do not split)

| Subfolder | Images | Labels | Notes |
|---|---|---|---|
| coco_train/ | 84,211 | 84,211 | Clean COCO person subset (img+label pairs match). |
| combined_train/ | 88,165 | 88,165 | **NOT clean**: only 3,954 img+label basenames match; 64,449 images have no paired label. Legacy merged folder — keep for completeness but do not treat as a ready-to-train pair set. |
| pets_train/ | 0 | 0 | Empty (PETS data lives in raw/; processed copy is empty here). |
| test_empty/ | 500 | ? (likely 0 or stub) | Empty-venue frames; used for TN/negative mining concept. Check label count before relying on it for training. |
| train/ | 3,954 | 3,954 | MOT16 training split (clean pair set, symlinked from raw/MOT16). |
| val/ | 1,362 | 1,362 | MOT16 validation split (clean pair set). |

**Upload recommendation:** upload the entire `data/annotated/` directory tree (images + labels + the 3 YAML files already in Tier 1) as a single HF dataset repo or as a folder inside a model repo. That preserves every subset and its messy overlaps, so future-you can pick the clean subset (coco_train, train, val) without loss.

---

## Tier 3 — DELETED locally (NOT uploaded; can be re-downloaded if ever needed)

### Rationale: large raw/public datasets that are either redundant with the annotated
###           corpus or re-downloadable. Deleting frees ~24 GB on a 94%-full C: drive.
###           All are recoverable from public sources (COCO, MOT16, PETS) if needed later.

### 3.1  Raw COCO 2017 (train + val + annotations)

| Path | Size | File count | What we did with it | Future need if deleted | Re-download source | Status |
|---|---|---|---|---|---|---|
| data/raw/COCO/ | **20.96 GB** | ~123,293 files (includes annotations JSONs, JPEG zips already extracted) | Source for COCO person images (15,000 sampled) + 5,000 background-only images used in Phase 4b extended training. Converted by coco_to_yolo.py into COCO person labels and background labels; results ended up in data/annotated/coco_train/ and data/annotated/labels/coco_train/. | **If deleted:** you can re-download COCO 2017 from the official source (cocodataset.org) or from Hugging Face datasets (e.g. `huggingface.co/datasets/coco`). The conversion scripts (Tier 1) are kept, so rebuilding annotated COCO labels is possible but would take time and bandwidth. **For Phase 5 inference: NOT needed.** **For future training from scratch using COCO backgrounds: needed again unless you already have the 5,000 background images preserved elsewhere.** | Public: cocodataset.org or HF datasets | DELETE locally after Tier 2 upload |

### 3.2  Raw MOT16 (train + test, with GT detections)

| Path | Size | File count | What we did with it | Future need if deleted | Re-download source | Status |
|---|---|---|---|---|---|---|
| data/raw/MOT16/ | ~1.9 GB | ~11,270 files | Source MOT16 frames + ground-truth detection files. MOT16 train frames were symlinked into data/annotated/images/train/ (3,954 images) and their GT labels converted to YOLO format in data/annotated/labels/train/. | **If deleted:** you lose the raw frames, but data/annotated/images/train/ + labels/train/ retain the 3,954 training pairs. For Phase 5 inference: NOT needed. For future retraining on MOT16: you'd re-download MOT16 (e.g. from MOTChallenge) and re-run the conversion symlinks. Low marginal cost to re-download (~1.9 GB). **Recommended to delete** since the processed version is in Tier 2. | MOTChallenge (motchallenge.net) or HF | DELETE locally after Tier 2 upload |

### 3.3  Raw PETS 2009

| Path | Size | File count | What we did with it | Future need if deleted | Re-download source | Status |
|---|---|---|---|---|---|---|
| data/raw/PETS2009/ | small (measured: 123 files, size not large) | ~123 files | PETS benchmark frames + annotations. Converted by pets_to_yolo.py; results were supposed to land in data/annotated/pets_train/ (which is currently empty — conversion may not have completed or was never run). | **If deleted:** PETS is a small public benchmark; re-downloadable if you ever need it. Current annotated/pets_train/ is empty, so deleting raw PETS has no immediate loss. **Low priority to keep.** | Public PETS benchmark site | DELETE locally (after confirming no dependency) |

### 3.4  Normalized MOT16 intermediate output

| Path | Size | File count | What we did with it | Future need if deleted | Status |
|---|---|---|---|---|---|
| data/normalized/ | **1.98 GB** | ~11,243 files | Intermediate output of MOT16 → YOLO normalization step (likely a copy/symlink tree of MOT16 images reorganized for training). Overlaps with data/raw/MOT16/ and data/annotated/images/train/. | **If deleted:** redundant with raw/MOT16 and annotated/train. Not needed for Phase 5 inference. **Safe to delete** — it was an intermediate artifact, not a final dataset. | N/A (regenerable from raw/MOT16) | DELETE locally |

---

## Model checkpoint bloat — DELETED locally (NOT datasets, but disk hogs)

### Rationale: epoch-by-epoch checkpoint files consume ~1 GB total across two training runs.
###           Only best.pt + last.pt are needed for inference / resume.

### 4.1  Phase 4 run checkpoint intermediates (train/ weights)

| Path pattern | Approx total | What we did with it | Future need | Status |
|---|---|---|---|---|
| models/finetuned/person_detector_v1/train/weights/epoch{0..16}.pt | ~150–250 MB total (17 files) | Per-epoch checkpoints from the earlier 29-epoch Phase 4 run. | **Not needed** for inference. Only best.pt (~6 MB) and last.pt are useful for potential resume. If you plan to resume this specific run later, keep last.pt; otherwise delete all epoch files. | DELETE epoch files; KEEP best.pt + last.pt |

### 4.2  Phase 4b extended run checkpoint intermediates (train_extended/ weights)

| Path pattern | Approx total | What we did with it | Future need | Status |
|---|---|---|---|---|
| models/finetuned/person_detector_v1/train_extended/weights/epoch{0..48}.pt | ~600 MB–1 GB total (49 files) | Per-epoch checkpoints from the 47-epoch extended run (the one with mAP50 ≈ 0.739). | **Not needed** for inference. best.pt is the converged model; last.pt is the resume fallback. If you might resume this run later (e.g. continue training on real venue data), keep last.pt. Otherwise delete all epoch files to free ~1 GB. | DELETE epoch files; KEEP best.pt + last.pt |

---

## Summary of deletions (local disk)

| Delete? | Path | Size saved | Notes |
|---|---|---|---|
| YES | data/raw/COCO/ | ~20.96 GB | Largest win; re-downloadable from COCO/HF if ever needed. |
| YES | data/raw/MOT16/ | ~1.9 GB | Redundant with annotated/train; re-downloadable from MOTChallenge. |
| YES | data/raw/PETS2009/ | small | Public benchmark; annotated/pets_train/ is currently empty. |
| YES | data/normalized/ | ~1.98 GB | Intermediate artifact; regenerable from raw/MOT16. |
| YES | models/.../train/weights/epoch{0..16}.pt | ~0.2 GB | Keep best.pt + last.pt only. |
| YES | models/.../train_extended/weights/epoch{0..48}.pt | ~0.8 GB | Keep best.pt + last.pt only. |
| **Total potential space freed** | | **~25–27 GB** | Before deleting, verify Tier 1 + Tier 2 are uploaded to HF. |

---

## What is NOT deleted (kept locally, regardless of HF upload)

- models/finetuned/person_detector_v1/train_extended/weights/best.pt (~24 MB) — primary detector for Phase 5.
- models/finetuned/person_detector_v1/train_extended/weights/last.pt (~24 MB) — resume fallback (keep unless you are sure).
- models/finetuned/person_detector_v1/train/weights/best.pt (~6 MB) — earlier run backup.
- models/pretrained/yolov8n.pt (~6 MB) — base weights for any future fine-tuning.
- config/ (both YAMLs).
- scripts/ (all .py files).
- data/annotated/dataset.yaml, dataset_coco.yaml, dataset_combined.yaml.
- data/annotated/images/ (full tree — Tier 2).
- data/annotated/labels/ (full tree — Tier 2).
- The CCTV test videos in dumps/video/ (1.mp4–9.mp4).
- All source code: src/, run_pipeline.py, verify_test.py, pyproject.toml, requirements.txt, etc.

---

## Future-workflow decision guide

### Phase 5 (current): Multi-Object Tracking Integration (BoT-SORT)
- **Requires:** fine-tuned best.pt (Tier 1), config/, scripts/, src/, test videos.
- **Does NOT require:** any raw datasets, annotated images/labels, or epoch checkpoints.
- **Action:** after Tier 2 upload, you can safely delete all Tier 3 + epoch bloat and still run Phase 5.

### Future: Fine-tuning on real venue data (when it arrives)
- **Requires:** base yolov8n.pt (Tier 1) OR an existing fine-tuned checkpoint to resume from (best.pt / last.pt from Tier 1), plus the new venue images + labels (to be collected), plus fine_tune.py (Tier 1).
- **Does NOT require:** COCO raw, MOT16 raw, PETS raw, normalized, or the old annotated corpus — unless you want to mix old data with new venue data.
- **If you want to keep the ability to re-mix old COCO backgrounds with new venue data:** keep the COCO-to-YOLO conversion script (Tier 1) and consider keeping the 5,000 background images (they may be in annotated/coco_train or test_empty — check which subfolder holds them before deleting). The COCO raw (20.96 GB) can be deleted because the *sampled* backgrounds are already in annotated/ (Tier 2).
- **If you never plan to reuse COCO/MOT16/PETS again:** deleting Tier 3 is zero-risk.

### Future: Training from scratch on a new dataset
- **Requires:** yolov8n.pt (Tier 1) as the base, a new dataset YAML + images + labels (to be created from your venue data), and fine_tune.py (Tier 1).
- **Does NOT require:** any of the old data. Tier 3 deletion is safe.
- **If you also want the option to re-create the old combined dataset exactly** (e.g. for reproducibility): keep Tier 2 (annotated/) on HF, and keep the conversion scripts (Tier 1). You would still need to re-download COCO raw + MOT16 raw to rebuild from scratch, but the scripts + YAMLs + annotated backup give you a recovery path.

### Recommendation for maximum future flexibility with minimum local footprint:
1. Upload Tier 1 (models + config + scripts + YAMLs, ~60 MB) to HF Hub.
2. Upload Tier 2 (data/annotated/ full tree, ~11 GB) to HF Hub as a backup dataset repo.
3. Verify both uploads landed correctly.
4. Delete all Tier 3 (raw COCO/MOT16/PETS, normalized) + epoch checkpoint bloat locally.
5. Keep Tier 1 locally as well (so Phase 5 runs without downloading from HF each time).

This leaves you with ~25–27 GB freed locally while retaining a full recoverable backup on HF and everything Phase 5 needs locally.