Detect-Objects

A trained object-detection product built from scratch for the vision-research portfolio. Give it any image and it returns bounding boxes + class labels + confidence scores across the 80 COCO classes.

Architecture: ResNet-50 + FPN → FA2/SDPA encoder → two-stage deformable decoder (RT-DETR/DINO-style). No anchors, no NMS.

Hardware target: 2× NVIDIA RTX 5090 32 GB (Blackwell sm_120) DDP.

Training data: full COCO 2017 (118 K train, 5 K val).

Product surface:infer.py CLI + demo_gradio.py web demo + safetensors / ONNX export for TensorRT deployment.

---

1. Headline capability (target)

"Detect-Objects: a from-scratch multi-scale deformable set-prediction detector trained on full COCO 2017 in ~5–7 days on 2× RTX 5090 32 GB, reaching ≈ 42 mAP on COCO val. Ships an inference CLI, a Gradio demo, and safetensors + ONNX export."

---

2. Quick start

cd Vision/detect-objects
pip install -r requirements.txt
# verify environment (RunPod 2× 5090)
bash scripts/env_check.sh
# run local smoke test (CPU OK)
bash scripts/smoke.sh
# run unit tests
PYTHONPATH=. python3 -m pytest tests/ -q

Inference on a single image

python3 infer.py \
--ckpt detect_coco_final.safetensors \
--in photo.jpg \
--out det.jpg \
--thr 0.5 \
--json detections.json

Gradio demo

python3 demo_gradio.py --ckpt detect_coco_final.safetensors --port 7860
# → open http://localhost:7860, upload an image, adjust the confidence slider

---

3. Training

Prepare data

python3 data/build_coco.py --out /workspace/data/coco

Launch full DDP run

CKPT_DIR=/workspace/runs/detect-objects bash scripts/launch_2x5090.sh

Training uses BF16 autocast, FP32 master weights, FP32 LayerNorm, EMA, gradient clipping, atomic safetensors checkpoints, and channels_last on the backbone. Resume is automatic from detect_latest.safetensors.

Evaluate

python3 -m training.eval_coco \
--ckpt /workspace/runs/detect-objects/detect_latest.safetensors

---

4. Results

Metric	Value
COCO mAP	pending 50-epoch run on 2× RTX 5090
mAP-50	pending
mAP-75	pending
mAP-S	pending
mAP-M	pending
mAP-L	pending
GPU util	target ≥ 95%
Wall time	~5–7 days on 2× RTX 5090 32 GB

Local verification (CPU)

bash scripts/smoke.sh # 2-ep, batch 2, 512², forward+backward, no NaN
PYTHONPATH=. python3 -m pytest tests/ -q

Smoke test and all unit tests pass locally:

• test_deformable_geom.py — 4-pt sampling stays inside image
• test_matcher.py — Hungarian finds optimal toy assignment
• test_fa2_encoder.py — encoder path under BF16 autocast + FP32 LayerNorm
• test_infer_cli.py — synthetic image → annotated JPEG + JSON

Sample detections will be added to this README after the GPU training run.

---

5. Architecture details

Component	Details	Params
Backbone	ResNet-50 from scratch + FPN (3 levels)	~25 M
Encoder	6 layers, d=256, 8 heads, FA2/SDPA self-attention	~10 M
Decoder	6 layers, 300 queries, multi-scale deformable cross-attention	~15 M
Total		~50 M

• Backbone:models/backbone_resnet.py — channels_last convs, no ImageNet pretrain.
• Deformable attention:models/deformable_attention.py — pure PyTorch grid_sample.
• Encoder/decoder:models/set_decoder.py — FA2 encoder + two-stage proposals + iterative refinement.
• Matching/loss:models/matcher.py + models/losses.py — Hungarian + focal/L1/GIoU.
• Training:training/train.py — hand-written DDP loop.

---

6. Export

# ship EMA-swapped safetensors
python3 export/to_safetensors.py --ckpt-dir /workspace/runs/detect-objects --out detect_coco_final.safetensors
# ONNX export (TensorRT-ready)
python3 export/to_onnx.py --out detect_coco_head.onnx

---

7. Project files

• AGENTS.md — developer reference and routing
• SKILLS.md — smoke/train/eval/ship/debug playbooks
• configs/coco_r50_detect_2x5090.yaml — single source of truth

---

Author: Atandra Bharati · Part of the CoreProjects vision-research portfolio.