DetSeg

Research done on object detection and segmentation during my PhD. Enjoy!

📀 Main results

Results on the 2017 COCO validation set. The inference FLOPs and FPS are measured on the first 100 images of the 2017 COCO validation set using an NVIDIA GeForce RTX 3060 Ti GPU.

Object Detection (FQDet)

Backbone	Head	Epochs	AP	Params	GFLOPs	FPS	Script	Log	Cp
R50+FPN	FQDet	12	43.3	33.9 M	99.0	20.9	script	log	cp
R50+TPN	FQDet	12	45.5	42.2 M	107.8	13.6	script	log	cp
R50+DefEnc-P3	FQDet	12	47.2	44.1 M	234.8	9.7	script	log	cp

Object Detection (FQDetV2)

Backbone	Head	Epochs	AP	Params	GFLOPs	FPS	Script	Log	Cp
R50+FPN	FQDetV2	12	47.0	37.9 M	117.4	17.7	script	log	cp
R50+DefEnc-P3	FQDetV2	12	50.8	48.1 M	256.1	15.5	script	log	cp
R50+DefEnc-P2	FQDetV2	12	51.7	48.4 M	747.0	6.8	script	log	cp
SwL+DefEnc-P3	FQDetV2	12	58.2	218.7 M	875.4	5.8	script	log	cp

Instance Segmentation (EffSeg)

Backbone	Head	Epochs	AP	Params	GFLOPs	FPS	Script	Log	Cp
R50+FPN	Mask R-CNN++	12	41.3	40.5 M	226.7	10.4	script	log	cp
R50+FPN	PointRend++	12	42.0	40.8 M	296.2	6.6	script	log	cp
R50+FPN	RefineMask++	12	42.7	44.2 M	455.1	6.3	script	log	cp
R50+FPN	EffSeg (ours)	12	42.4	41.8 M	262.6	7.5	script	log	cp

Panoptic Segmentation (EffSeg)

Backbone	Head	Epochs	PQ	Params	GFLOPs	FPS	Script	Log	Cp
R50+FPN	Mask R-CNN++	12	45.8	40.6 M	218.6	9.9	script	log	cp
R50+FPN	PointRend++	12	47.0	40.9 M	289.7	6.3	script	log	cp
R50+FPN	RefineMask++	12	47.2	44.2 M	433.2	6.3	script	log	cp
R50+FPN	EffSeg (ours)	12	47.0	41.8 M	262.6	6.7	script	log	cp

📃 Papers and thesis

• Trident Pyramid Networks for Object Detection by Cédric Picron and Tinne Tuytelaars.
• FQDet: Fast-converging Query-based Detector by Cédric Picron, Punarjay Chakravarty, and Tinne Tuytelaars.
• EffSeg: Efficient Fine-Grained Instance Segmentation using Structure-Preserving Sparsity by Cédric Picron, and Tinne Tuytelaars.
• Designing High-Performing Networks for Multi-Scale Computer Vision by Cédric Picron (PhD thesis).

🛠️ Installation

• Environment:

  1. Install the conda package and environment management system if not already done.
  2. Execute source setup_env.sh.

• Data preparation:

  1. Download the desired datasets.
  2. Modify the paths in setup_data.sh to point to your installation directories.
  3. Execute source setup_data.sh.

🌱 Usage

• Training: Execute python main.py with the desired command-line arguments. Some example training scripts, which were used to obtain the results from above, are found in the scripts directory.

• Evalutation: Execute python main.py --eval --eval_task $TASK with the desired command-line arguments, with $TASK chosen from:

  1. analysis: Analyze the computional cost of the given model.
  2. comparison: Compare the results from two different models.
  3. performance: Compute the model performance on the desired benchmark.
  4. profile: Profile the given model.
  5. tide: Perform TIDE analysis of given model.
  6. visualize: Visualize the model predictions.